The recent three-hour podcast featuring computer scientist Lex Fridman and Prime Minister Narendra Modi gained widespread attention. Streamed in Hindi, English and even Russian, it has been called the “best dubbing” to date by many. The hyperrealistic AI translations, nearly indistinguishable from the original, were made possible by ElevenLabs, a three-year-old AI startup. 

The technology was also deployed to translate Lex’s interview with Ukrainian President Volodymyr Zelenskyy in Kyiv in English, Ukrainian, and Russian.

While this is not the first time PM Modi has used AI for translation, it was a crucial opportunity for ElevenLabs to showcase its technology by helping world leaders reach a global audience in multiple languages. 

Siddharth Srinivasan, who is heading ElevenLabs in India, told AIM that India is ElevenLabs’ biggest market yet, and the company is actively expanding its team in the country to build the future of voice AI.  

He believes the opportunity is immense for a country like India, given its diversity in languages and accessibility needs. 

Srinivasan revealed that ElevenLabs is nearing a team of 10 employees in the country, currently focusing on business-related roles, and has plans for further expansion.

This aligns with the trend of Western labs and startups establishing their presence in India. OpenAI and Perplexity are also reportedly preparing to expand in the country soon.

Last month, the Poland-based startup closed another funding round, raising $108 million at a valuation of $3.3 billion. ElevenLabs went from a weekend project to a startup in April 2022, when founders Mati Staniszewski and Piotr Dabkowski set out to solve poor dubbing with lifelike voice synthesis.

Relevance of Voice UI 

While there is significant competition in the voice AI space, including traditional tech companies, ElevenLabs’ differentiation in the Indian market comes from catering to different accents and languages accurately.

ElevenLabs offers several distinct models for its AI audio technology, specifically tailored for different use cases, such as text-to-speech (TTS) and speech-to-text (STT). 

“We have the fastest model in the world in Indic languages, which is…speech-to-text language,” Srinivasan added. 

As of now, the company supports 11 Indian languages, with plans to expand further. However, Srinivasan noted, “If you get about eight to 11 languages on both sides, [you cover] 70% of India.”

Limited digital data for many Indian languages remains an obvious challenge. ElevenLabs is addressing this by building datasets through strategic partnerships and community involvement. 

Overall, the company’s multilingual AI model supports 29 languages, providing highly realistic, emotional voices in each. Moreover, their conversational AI model supports 32 languages, enabling natural, real-time conversations. They’ve also developed an advanced speech-to-text model in 99 languages, including 11 Indian languages.

Growing Creator Economy in India 

ElevenLabs has effectively productised its technology to cater to content creators and developers. Their API is user-friendly, allowing large-scale integration, particularly appealing to developers and businesses looking for scalable solutions. 

Podcasters generally are a huge target audience for the company. 

In addition to Fridman, Indian tech podcaster Varun Mayya employs ElevenLabs’ technologies for his own brand and other companies. The company has also partnered with neuroscientist and podcaster Andrew Huberman to dub content for his Huberman Lab podcast into Hindi and Spanish. Their partnerships with Spotify help in producing AI-narrated audiobooks. 

Srinivasan mentioned how Star Sports uses its technology to dub Steve Smith’s voice in Hindi and Tamil, and localising cricket content for a broader audience. 

In terms of new features and partnerships, last year ElevenLabs launched GenFM (a feature similar to Google’s NotebookLM) for multispeaker podcasts and partnered with Pocket FM to help writers turn stories into audio. 

In addition to the content creation market, industries of focus in India include customer support and education. 

Recently, the company also hosted a hackathon in Bengaluru as part of its global tour, with India recording the highest number of registrations. 

The showcased projects included an AI agent for emotional support, an AI video framework with designated roles for content creation, and a voice AI solution for rural India to enhance accessibility in areas where chatbots are limited.

Future of AI is Not Unipolar 

Srinivasan, who comes with a decade of experience at YouTube, understands the creator ecosystem in the country. He believes the future of AI will not be limited to just one interface – although voice will be a massive pillar of it. 

He pointed to a consistent trend in media and technology towards multimodal interactions, and even going forward people will engage with AI across voice, text, video, and visuals.

“Even before the rise of generative AI, people were engaging with technologies like Alexa, Google Assistant, and Siri,” he said.

But, he is confident that voice will remain the most natural interface, as it continues to be the most basic and widely used form of communication, deeply embedded in both consumer and business interactions.

Srinivasan also acknowledged the risks of deepfakes and emphasised that ElevenLabs uses moderation, permission for consent, and traceability to prevent misuse.

At the recent Arkam Annual Meet 2025, themed ‘The Great Unlock India in 2035’, entrepreneur and Infosys co-founder Nandan Nilekani sparked a cheer from the crowd when he said by 2035 there would be a million startups. He quickly followed up with, “Just saying. You guys have to make it happen.”   

Nilekani, who outlined the path to reach an $8 trillion economy in the next decade, believes India’s startup ecosystem will play a crucial role in achieving this goal. 

“When the economy is growing at 8%, startups will grow at 20%, reaching a million startups by 2035,” Nilekani said. He stressed that the surge in entrepreneurship, combined with the nation’s Digital Public Infrastructure (DPI) and AI adoption, is expected to fuel unprecedented economic acceleration.

1 Million Startups

Nandan Nilekani at Arkam Annual Meet 2025. Source: AIM

The rapid formalisation of businesses, access to digital markets, and the increasing availability of capital are setting the stage for India’s next wave of startup-driven economic growth. Nilekani highlighted how platforms like Meesho are enabling first-time sellers to tap into a national customer base. 

“Most of the people on Meesho are selling for the first time on an online platform. Most of the buyers of Meesho are buying for the first time on an online platform. So, platforms like Meesho will actually allow small guys to sell to a national market.”

India’s DPI has laid the groundwork for rapid startup expansion by reducing barriers to entry, enhancing financial inclusion, and simplifying compliance. With Aadhaar, UPI, and Account Aggregators (AA), startups now have the tools to scale faster than ever before.

Nilekani pointed out that India’s formalisation rate remains low, but technology is changing that. “The technology is now available for small guys to modernise themselves. In some cases, they are forced to modernise because they have to pay taxes, they have to use GST. A combination of push and pull will bring millions of these companies into the formal sector in the next decade,” he said. 

He emphasised that access to digital identity, easy credit through account aggregators, and seamless taxation systems will enable a thriving digital-first economy.

India has already witnessed massive fintech adoption, with UPI leading the way. “PhonePe has 350 million active users. It has 48% of the payment market,” he stated, underscoring how digital payments are now an integral part of economic transactions.

The next step, he believes, is making benefits and credentials portable, ensuring economic mobility for workers and entrepreneurs alike. “If I can create a truly national market for benefits, so I can go anywhere and take my benefits along with me, that’s when we truly unlock population-scale formalisation.”

The AI Combo

Source: AIM

Alongside DPI, AI is emerging as a critical enabler of startup-driven growth. With AI-native Gen Z entrepreneurs entering the workforce, the way businesses operate is undergoing a shift. 

“We know that half of India’s workforce will be AI-made. We will have 380 million Gen Z people. These people haven’t seen the world without a phone. They will be the first to use all this AI,” Nilekani said.

Nilekani goes on to explain that startups using AI-driven automation, predictive analytics, and personalised services will have a significant competitive advantage. India’s push for AI-driven multilingual interfaces will further democratise access to technology, enabling more entrepreneurs from tier-2 and tier-3 cities to participate in the economy.

Looking Ahead

In addition to highlighting startups that are solving unique problems, Nilekani predicts that the next mass-scale economic transformation will come from energy trading, similar to what UPI did for financial transactions. He envisions millions of small energy producers, ranging from households with solar rooftops to EV battery owners, participating in an open energy marketplace.

“Every home will be an energy producer because they have rooftops. Every home will be an energy storer because they have an EV battery. So, every home is a producer of energy, seller of energy, buyer of energy…You’re going to buy and sell energy, like UPI,” he explained.

With the convergence of one million startups, DPI, AI, and energy decentralisation, India is on the cusp of a historic economic transformation. As Nilekani predicts, this could probably be the path to both an $8 trillion economy and entrepreneurship explosion.  

Conversations on AI startups building wrappers on foundational models have been ongoing for the past few years. The debate about the value of startups centred on these wrappers is more complicated than it initially appears, with no conclusive verdict reached so far. 

In the recent Google Cloud report titled ‘Future of AI: Perspectives for Startups 2025’, various startup founders and venture capitalists (VCs) shared their insights on AI trends and predictions. One notable perspective came from David Friedberg, entrepreneur and CEO of Ohalo Genetics, who believes that it’s not enough to just be an LLM wrapper. 

“If you’re just an LLM wrapper, it’s going to be hard to build a sustainable business—you’re likely

going to get commoditised away,” said Friedberg, who believes that businesses need a mechanism for value creation that sustains an initial competitive advantage through ongoing enhancements. 

“This will typically come from proprietary data generation, which is used to continuously improve model performance, or network effects that lock-in access to data or customers,” he said. 

He believes that startups need to be agile and adaptable, but at the same time, they must also deliver unique value. 

While Friedberg’s thoughts may be coming from an investor’s point of view who is probably looking for long-term value from startups, a number of AI startups springing up are mostly AI wrappers. 

Innovative Solutions Matter

“Most people are so upset when they use the word wrapper. It feels so derogatory. You feel like an insulter or something,” Perplexity co-founder and CEO, Aravind Srinivas said in an interview earlier. 

Srinivas highlighted how people perceive the term ‘wrapper’ today, and suggested that in many ways, everyone is a wrapper. Notably, Perplexity integrates multiple LLMs alongside their own model to deliver more comprehensive results. 

Recently, Will Poole, managing partner and co-founder of Capria Ventures, shared a similar viewpoint. During an exclusive video interaction with AIM, Poole noted that while some might view wrappers negatively, others draw parallels to how SaaS is essentially built as a wrapper around databases, despite SaaS having an entire industry built around it.

Poole stated that instead of wrapper’s perception as derogatory, the focus is on whether true, deep innovation has been applied and whether it generates valuable business data that can create a competitive moat, enabling long-term sustainability.

Considering how a number of Indian startups cater to various markets, the ability to customise and build upon existing LLMs presents a solution to complex problems. In fact, VCs are looking to invest in Indian AI startups that offer nuanced solutions that big tech companies cannot deliver. 

Citing 5C Network as an example, Poole explained how the medical AI startup is helping transform radiology departments by running diagnostics via AI. The key differentiator is their access to high-quality and diverse medical data, which ultimately serves as a moat for them. It also adds to the reason why Indian companies would prefer to work with home-grown startups rather than big tech companies who have built their proprietary models. 

Not Shying From Wrappers

Leading VC firm Accel, which is aggressively investing in emerging AI startups in India, with a recent announcement of $650 million in early-stage funds to boost AI and tech for Bharat, had earlier shared a similar sentiment on investing in wrapper companies. 

“The majority of people can start with a wrapper and then, over a period of time, build the complexity of having their own model. You don’t need to do it from day one,” Prayank Swaroop, partner at Accel, earlier told AIM. 

Swaroop emphasised that Accel has no inhibition in investing in wrapper-based AI companies, as long as the startup can prove its ability to find customers by building GPT or AI wrappers on other products. However, he said that for a research-led foundational model, it is crucial to stand out, and simply creating a GPT wrapper does not qualify as a new innovation.

Source: X

Recently, a Chinese general purpose AI agent Manus has been taking the world by storm for the last few days. A general agent can independently plan, execute, and produce complete results by browsing websites in real time and handling multiple data types for processing and generation. What is interesting is that Manus is an AI wrapper.

Manus was built on top of Anthropic’s Claude Sonnet model and other tools such as Browser Use. While this has sparked significant discussions, the reality is that people continue to build them, with the ultimate measure being the value addition they provide. 

Notably, some startups built around foundational models from big tech companies have received impressive valuations. One such example is Harvey AI, a generative AI platform designed for legal professionals, which is built on OpenAI’s GPT models and backed by OpenAI.

The startup raised a total of $580 million in funding, including a recent $300 million Series D round, which valued the company at $3 billion. 

While the rise of AI wrapper startups reflects market sentiment, their adoption is a testament to the fact that wrappers are likely here to stay. Perhaps being an LLM wrapper isn’t as bad as it once seemed after all.

IIT Madras professor Balaraman Ravindran has cast doubt on the government’s ambitious plan to develop an indigenous large language model (LLM) under the IndiaAI mission within six months. 

In an exclusive conversation with AIM, Ravindran indicated that the timeline may be too short to build a top-notch model and doubted whether the initiative can truly position India on the global AI stage.

“I think six months is too aggressive a timeline for us to really build super capable models. What we are probably going to get are right or decent models; we are not going to shake the world,” he said.

This comes in the backdrop of Union minister Ashwini Vaishnaw expecting India’s LLM to be ready within the next ten months. The government allocated ₹2,000 crore in the Union Budget 2025 for the IndiaAI mission and has so far received 67 proposals to develop indigenous AI foundational models, including 22 LLMs, for India’s diverse linguistic and cultural landscape.

Although startups are still uncertain about the timeline, the Ministry of Electronics and Information Technology (MeitY) said it will continue accepting proposals till the 15th of each month for the next six months, or until it receives a sufficient number—whichever comes first.

The government on Thursday launched AI Kosha and Compute Portal under the IndiaAI Mission. These initiatives are part of the IndiaAI Mission to provide startups and researchers with access to datasets and high-performance computing resources for artificial intelligence development.

Taking academia as the reference point, Ravindran, who once taught Perplexity CEO Aravind Srinivas, stressed how short six months are. “[In that time] one semester happens—assignments and exams happen. Unless the research has already been happening, you can’t expect an LLM in six months.”

Meanwhile, IIT Madras, in partnership with a startup, has also sent a proposal for the IndiaAI mission. “Whoever submits a proposal has to set up as a company because six months of delivering something requires you to have a startup mindset,” the professor said. 

He opined that the government should have invested more in academia. “You cannot really do cutting-edge automotive research in academia alone,” he said, adding that it has come to a point where if one wants to do cutting-edge AI research, one needs to partner with the industry. 

Ravindran acknowledged that the current plan of the IndiaAI mission is better than the original plan of buying a lot of GPUs directly from NVIDIA. “Now, at least, we have these cloud vendors who can hopefully provide more subsidised things,” he said. 

The IndiaAI mission will feature 18,000+ GPUs through public-private partnerships with players like Jio Platforms, NxtGen Data Center, Locuz Enterprise, E2E Networks, CtrlS DataCenters, CMS Computers, Orient Technologies, Tata Communications, Vensysco, and Yotta Data Services.

AIM found that AI startups such as Sarvam AI, Krutrim, CoRover.ai, TurboML, LossFunk and IIIT Hyderabad had also applied under the proposal.

Notably, Soket AI recently launched Project EKA, which seeks to unite AI researchers, engineers, and institutions across the country to develop multilingual, high-efficiency AI models that cater to India’s needs while competing globally.

Similarly, TurboML, a San Francisco-based AI company, has launched a new initiative to bring together Indian-origin AI researchers from around the world. The goal is to develop foundation models to support the country’s future in AI. 

“We’re collaborating with the best minds to build foundation models that will power India’s AI future”, wrote TurboML CEO Siddharth Bhatia on X. In his post, he mentioned discussing the initiative in a meeting with IT minister Ashwini Vaishnaw. 

Bhatia told AIM that his company plans to build an LLM under $12 million. 

Indian SaaS company Zoho has also entered the race.  In a recent interview, group chief executive officer Shailesh Davey said the company plans to release two foundational models by the end of the year. Zoho is also developing foundational models for Indian languages but has not set a timeline on this yet. 

More Investment and Compute Needed

Ravindran said that another advantage of the mission is that it will create a pool of trained engineers who can work with more researchers in the future and develop better solutions. However, he is still not convinced that India can build a DeepSeek-like model in six months but is hopeful that something fruitful can come in the next year and a half.  “If you build DeepSeek in a year and a half, it has to be something very, very different,” he said.

The professor pointed out that India has a dearth of corporations investing money in fundamental research. “They are happy to do applied research,” he said.

He explained that China’s DeepSeek didn’t come out of nowhere and that research had been going on for a long period. “As far as China is concerned, there were at least 5-6 companies, maybe 10, that were at the level of DeepSeek, and they could just launch anytime,” Ravindran said.

He even suggested that the timing of DeepSeek’s release seemed strategic. “They stayed silent until the US announced the Stargate initiative.”

This, he said, was a clear response to the developments in the US. “Stargate said it needed $500 billion, so the Chinese countered with, ‘We can do it in $4 or $6 billion’,” he said, suggesting a political angle to the whole affair.

“It’s not like DeepSeek V3 just came out of the blue. If you look at a lot of things that they used in DeepSeek V3, many of the components are there in DeepSeek V2. They added RL fine-tuning instead of supervised fine-tuning in the end,” Ravindran said. 

He added that American companies were trying to create a moat around themselves by saying that you need a lot of money to get to the same level as us, and that’s perhaps one of the reasons they didn’t want to explore cheaper alternatives.

Ravindran said that when discussing India needing more compute, he means reaching a sustainable level, not an ‘Elon Musk-level’ of compute. Notably, Musk’s xAI trained Grok 3 on 200,000 GPUs. The IIT professor believes India needs to look beyond Transformer-based autoregressive training and even reinforcement learning-based fine-tuning. “I think they will soon be hitting a ceiling. People who worked on symbolic AI were way ahead of their time in how they thought about intelligence,” he concluded.

Food and grocery delivery platform Swiggy handles billions of orders annually across approximately 700 cities. Managing such a scale requires real-time coordination between restaurants, delivery partners, and customers. This is where data streaming platform Confluent steps in, supporting Swiggy’s massive data flow for seamless order processing.

As the exclusive data streaming partner, Confluent has been supporting Swiggy for over three years now. 

At its core, Confluent’s platform is built on Apache Kafka, facilitating real-time data streaming. Notably, Confluent was founded in 2014 by Jay Kreps, Neha Narkhede, and Jun Rao, the original creators of Apache Kafka.  

“If Confluent is down, Swiggy is down,” Rohit Vyas, country presales leader of India at Confluent, told AIM in an exclusive interaction. He emphasised the platform’s role in Swiggy’s infrastructure.

Scale and Demand Spikes

Data streaming platform Confluent has been in the space for over a decade, catering to more than 5,000 enterprises. 

“We are the second fastest software company to reach a billion dollars in revenue,” said Rubal Sahni, area vice president and country manager at Confluent, to AIM.  

Catering to different domains, Confluent recently partnered with Jio Platforms to integrate its data streaming platform with JioCloud Services. Another major customer is Swiggy, where real-time data processing is mission-critical.

Swiggy has processed over 3 billion orders, requiring a system that can scale dynamically. Initially, Swiggy relied on open-source Apache Kafka for its data streaming needs. However, as order volumes surged, maintaining an in-house Kafka setup became resource-intensive for Swiggy’s engineering teams. 

This led to the transition to Confluent’s fully managed service, which provides real-time data streaming while reducing operational overhead. This move not only enabled Swiggy to streamline operations, and handle order surges efficiently, but also optimise resource allocation during high-demand periods such as festivals.

“So when they scaled up in a hyper scaling manner and their customer base began to bulge on, they realised that babysitting open source Kafka was draining a lot of their engineering talent. Then they approached Confluent,” Vyas explained. “What we enable for them is a multitude and a multi-dimensional aspect.” 

Real-Time Data Processing

Confluent plays a critical role in enabling Swiggy to track order progress in real time, ensuring transparency at every stage. By ingesting vast amounts of data and processing it within milliseconds, the platform updates order statuses, estimated delivery times, and driver locations without delays. “If Confluent is down, apparently, Swiggy is down,” Vyas said. 

From the moment a customer places an order, Confluent’s data streaming platform ensures that every step, from restaurant confirmation to assigning a delivery partner, happens seamlessly.

This entire complex system of fulfillment is happening over Confluent, making real-time updates possible, Vyas explained. The platform dynamically maps delivery partners, estimates preparation and packing times, and continuously syncs this information with Swiggy’s app. 

Beyond scaling, ensuring security and governance is another challenge. Swiggy processes massive volumes of customer data, making compliance and security a top priority. Confluent provides built-in governance and monitoring tools, helping Swiggy maintain regulatory compliance while securing sensitive information.

Beyond logistics, Swiggy utilises AI-driven analytics powered by real-time data streaming. Predictive algorithms help optimise delivery routes and inventory management. “Swiggy Instamart tracks high-demand items in specific locations, ensuring better stock availability in dark stores,” Vyas further said.

The Future Ahead

As Confluent continues to expand its presence in India, the company is focusing on scaling its industry reach and strengthening partnerships. “We have scaled our teams in India. We have opened newer industry verticals. We have invested in the public sector,” Sahni said. The recent partnership with Jio Platforms aims to support public sector organisations with infrastructure for GenAI applications. 

Beyond public sector expansion, Confluent is also increasing its footprint across industries such as healthcare, manufacturing, and tier-two markets. “We are adding more resources to cater to newer industry verticals, not just metro cities but also down south and up north,” Sahni added. To further strengthen its presence, Confluent is working closely with hyperscalers like AWS, Azure, and Google Cloud while expanding partnerships with enterprises looking for scalable data solutions.

The company will continue to cater to markets of different sizes, from startups to large-scale enterprises. “We have various offerings depending on the size of the customer and nature of use cases,” Sahni said. 

India’s AI landscape is witnessing rapid growth, however, access to early-stage capital remains challenging. While large venture capital (VC) firms and global investors often back late-stage AI startups, early-stage funding is increasingly scarce. 

This is precisely where micro VCs step in and fill the funding gap. 

Shift to MicroVCs

The Indus Valley Annual Report 2025 (Blume Report) highlights a clear trend: larger seed rounds or ‘mango seeds’ (>$3 million) now account for half of total early-stage funding, while sub-$1 million rounds have significantly declined. 

This shift has made it harder for first-time AI founders to secure capital, opening up opportunities for micro VCs to step in.

“India’s innovation ecosystem depends heavily on micro VCs, particularly as the window for creating AI-first product-tech companies gets smaller,” said Ranjeet Shetye, venture partner at YourNest Venture Capital, in an exclusive interview with AIM. 

“MicroVCs can support more early-stage startups by writing smaller checks, which spreads the risk and increases the possibility of power-law outcomes, where a few breakthrough successes generate enormous returns,” he said. 

The Rise of MicroVCs for AI

Micro VCs, which are smaller, highly specialised funds that invest in pre-seed and seed-stage AI startups, differ from traditional venture firms in their approach. According to the report, over 100 micro VCs in India typically invest between $100,000 and $500,000 at the seed or pre-seed stage. 

Unlike larger VCs that prioritise proven founders and market traction, micro VCs focus on high-risk, high-reward opportunities, especially in deep tech and AI.

Another key advantage of micro VCs is their domain expertise. Unlike generic seed funds, micro VCs often specialise in AI, SaaS, or deep tech, providing targeted support beyond capital. This includes technical mentorship, AI model optimisation, and go-to-market strategies tailored for AI startups. 

Their ability to invest at valuations of $1 million to $8 million allows them to take early bets on disruptive AI technologies even before mainstream investors take notice.

Source: Indus Valley Annual Report 2025

MicroVCs Driving India’s Next AI Unicorn

AI-first startups require substantial capital to train and refine LLMs. However, capital efficiency remains key to early growth. Shetye emphasises that micro VCs play a critical role in shaping India’s next AI giant.

“Founders can achieve product-market fit, iterate quickly, and drive sales-led growth by supporting AI-first product development and customer success models early on,” Shetye explains. 

“Reaching milestones like $1 million ARR faster is enabled by this lean, capital-efficient strategy, which also speeds up follow-on funding rounds.”

This model has already shown promise. Startups backed by seed funds typically take longer to raise Series A rounds compared to those funded by multistage VCs. 

However, with strategic backing from micro VCs, AI startups can optimise early-stage growth while keeping burn rates low. This increases their chances of securing larger investments down the line, fueling their journey to unicorn status.

Interestingly, India’s push for AI sovereignty has led to increased government support for AI startups. The ₹20 billion ($240 million) AI Sovereignty Fund, announced in the latest Budget, aims to spur homegrown LLM development and AI research. 

Additionally, Indian startups now have access to 18,000 GPUs at 40% below market rates, providing a significant cost advantage in AI training and deployment.

This policy shift aligns well with micro VC-backed startups that prioritise lean AI model development. While US and Chinese AI firms raise billions for model training, Indian AI startups are adopting frugal innovation strategies, making breakthroughs with smaller budgets.

Not That Simple 

Despite their benefits, micro VCs face challenges in scaling their influence. One of their biggest hurdles is the struggle to raise capital themselves, making it difficult to maintain consistent funding pipelines. Additionally, they operate in a high-risk environment where a large percentage of portfolio companies may not survive past Series A.

Another challenge is the scarcity of specialised AI talent. While India has a strong engineering base, it lacks the deep AI research ecosystem seen in the US and China. Many AI founders prefer to move abroad for better funding and infrastructure. 

To counter this, micro VCs are now collaborating with academic institutions such as IITs and IISc to nurture AI talent at home.

As India aims to build its own DeepSeek-like AI models, with the recent being Soket AI Labs’ Project EKA that is inviting researchers and developers to build sovereign AI models, micro VCs could play a pivotal role in funding the next generation of AI leaders. 

The Indian space sector, once dominated by government-led initiatives, is seeing a surge in private players, opening up career trajectories that didn’t exist. Along with that, the decade-old norm of space engineers studying in India, gaining experience at ISRO, and then moving abroad for better opportunities is also fading away. 

Instead, more Indian engineers are choosing to stay and build in India. 

One of the catalysts for this shift has been India’s evolving space policy. “If you recall, in 2020, Nirmala Sitharaman announced approximately ₹20 lakh crore as part of the COVID package. A portion of that was dedicated to opening up new sectors for increased employment and investments—space was one of them,” said Yashas Karanam, co-founder and COO of Bellatrix Aerospace, in an exclusive interaction with AIM. 

The announcement triggered confidence, leading to a wave of private space ventures, which in turn created jobs and opportunities that could compete with global offers.

Giving Engineers a Reason to Stay

India boasts over 190 space startups, a significant jump from just a handful in 2015. Karanam said that companies are not only building satellite technologies but also innovating in propulsion, launch systems, and AI-driven space navigation.

“In 2012, there was probably one company. In 2015, there were three. By 2017-18, it was up to 10. But after 2019-2020, the acceleration has been significant,” said Karanam.

The investments tell the same story: Indian private space startups have collectively raised over $350 million in funding in the past five years, demonstrating both investor confidence and a growing market.

This expansion is keeping engineers engaged in the country. “Now the opportunities are getting better here in India,” Karanam explained. “There are really exciting challenges to solve, and in terms of the trade-off, people went abroad because of better projects and financial incentives. But if you are paid well here, you could visit those countries while still building in India.”

In terms of talent, Karanam believes India has the right set of people. “I feel we are really good with what it takes to build a space company or space products because the engineers we have got have been really good. A lot of times, training or onboarding onto particular technologies would be required, but everybody is a quick learner as far as we have seen,” he said. 

ISRO to Private Startups

Karanam observes an emerging trend within this space boom, which is the migration of former ISRO scientists to the private sector. Unlike earlier, where ISRO was seen as the only credible player, today’s engineers, many of who have worked on India’s high-profile missions, are stepping out to start their own companies.

“We have seen more people quitting their ISRO jobs and starting up on their own because they now know that the government is supporting this sector,” Karanam said. “India also realised that despite being one of the top five global spacefaring nations, we had less than 2% of the global commercial market share. That was because only ISRO was doing it.”

This realisation has fueled a policy shift that encourages private participation, resulting in companies securing contracts directly from ISRO and IN-SPACe (India’s nodal space authorisation body). The industry, once restricted to working under ISRO’s guidance, is now developing independent solutions for both domestic and international clients. 

Bengaluru-based aerospace manufacturer Bellatrix Aerospace, which specialises in developing in-space propulsion systems and small satellite manufacturing, was founded in 2015 by Rohan M Ganapathy and Karanam. Since its inception, Bellatrix has raised a total of $11.3 million over four funding rounds. 

Some of its notable investors include Inflexor, Pavestone, StartupXseed, GrowX, BASF, and others. Bollywood actress Deepika Padukone has also backed Bellatrix. 

This Indian private space startup has developed two innovative satellite propulsion systems: a water-based Microwave Plasma Thruster and India’s first private Hall Effect Thruster, positioning the country as a leader in innovative propulsion technology.

‘Make in India’ for Space Hardware 

Despite the enthusiasm, India’s space startups face a significant hurdle: building hardware at scale. While software-based AI and deep-tech companies can scale quickly with limited capital, space hardware requires heavy upfront investment, regulatory approvals, and access to cutting-edge materials.

Karanam elaborated on the challenges of building space components domestically. “If I were to build a propulsion system, I could buy a valve from a company that makes valves, buy propellant from a propellant manufacturer, and buy a catalyst from Shell or SABIC. 

“But if all my suppliers are outside India, and I have to import everything, i.e. paying space shipping and customs duty, it does not make sense in the long term,” he said.

As a result, companies are increasingly choosing to manufacture components in-house. “We were blessed with a really capable team, and we had several retired ISRO scientists advising us. So, we took the bold decision to build even materials in-house,” Karanam added.

While building space hardware is a challenge, funding for hardware development is a major roadblock. Most investors, accustomed to rapid returns from SaaS and consumer-tech startups, hesitate to back capital-intensive deep-tech ventures.

“There is a funny saying: ‘The space industry is a sinkhole for money.’ The more you deploy, the more it goes into capital and R&D,” Karanam said, reflecting on investor scepticism. Unlike software, which can show quick user growth and revenues, hardware requires patience, long development cycles, and extensive testing.

Amblyopia, often referred to as ‘lazy eye’, is a prevalent yet frequently overlooked vision disorder that affects 1-5% of the global population. Its prevalence has reached up to 5% in India. While conventional therapies exist and have proven effective, they come with numerous challenges and are often met with resistance, especially from young children. A Bengaluru-based health-tech startup is using immersive virtual reality to address this issue. 

Founded in 2023, NeuraSim uses advanced VR-based simulations to rewire the brain’s visual processing. It offers a breakthrough alternative to traditional treatments for amblyopia, such as patching therapy. 

“The conventional therapy is called patching, where we put a patch over the good eye and let the brain be forced to use the weaker eye. This takes nine months, and children often drop out within two to three months. In contrast, our immersive therapy improves vision starting from 20 days of therapy, with results in 45 days,” said Ramesh S Ve, founder and CEO of NeuraSim, in an exclusive interaction with AIM. 

VR-Based Vision Therapy

Amblyopia usually occurs in childhood when one eye is significantly weaker than the other. This leads the brain to suppress input from the weaker eye. If left untreated, it can result in permanent vision impairment. 

NeuraSim’s VR therapy offers a non-invasive, engaging, and scientifically validated approach. By immersing patients in controlled VR environments, the treatment stimulates the weaker eye while simultaneously ensuring the brain integrates inputs from both eyes. 

“We need to dissociate one eye as being weak and the other as being strong. The brain is conditioned to only use one eye, and we have to recondition it and rewire it. Virtual reality, by design, brings in that facility,” Ramesh said. 

NeuraSim VR headset

The therapy consists of immersive VR simulations customised for each patient’s needs. The technology has been validated through clinical testing, and the results are promising. 

“To our surprise, within 45 sessions, from near blindness in the weak eye, a seven-year-old child gained almost 600% improvement. Today, that child has stable vision which she wouldn’t otherwise have had,” Ramesh shared.

This form of treatment also addresses the shortage of ophthalmologists and optometrists in the country. “For India’s 1.4 billion population, we just have around 30,000 ophthalmologists and 25,000-30,000 optometrists. Even if we include opticians and other support staff, the total workforce is just about one lakh. This is nowhere near enough to provide adequate eye care for the growing population,” Ramesh emphasised. 

Software and Hardware Combo 

NeuraSim, founded in 2023 and headquartered in Bengaluru, has its research and development incubated at the Manipal Bioincubator. The company initially received funding from a Biotechnology Industry Research Assistance Council (BIRAC) grant and later secured a €1 million Indo-European research grant to develop its VR technology.

NeuraSim’s VR solution is built on years of research and clinical trials. Ramesh, who has a background in public health and vision science, worked with leading institutions like Johns Hopkins University and Manipal Academy of Higher Education before founding NeuraSim with Girish Somvanshi. His extensive experience in epidemiological studies and telehealth inspired the creation of the startup. 

Despite being a relatively young startup and even recently getting featured in Shark Tank India, where they failed to receive funding, NeuraSim has already gained traction among India’s top eye care institutions and received a number of inquiries following the airing of that episode. 

Aravind Eye Care, an ophthalmology institution known worldwide, was among its first paying customers. “We wanted to validate if there was a real business case. Today, we have over 20 customers, including repeat clients from the Aravind chain,” Ramesh said.

NeuraSim offers two pricing models for its VR-based amblyopia therapy. One of the models includes a monitor, mini-computer, and VR device and is priced at ₹4.5 lakh. For this variant, therapy is administered by a clinician. Meanwhile, the second model is available for rent at ₹12,000 per month. “That comes to around ₹200 per session, making it more accessible for parents who would otherwise spend ₹400 on travel for clinic visits,” he further said.

While VR-based vision therapy shows great promise, the road to adoption doesn’t come without its challenges. One of the primary hurdles was the availability of suitable VR hardware. 

Initially, NeuraSim relied on imported devices like Pico, which later became difficult to procure due to geopolitical issues. This pushed the company to partner with Pune-based startup Question What’s Real (QWR) to manufacture VR headsets locally. 

Hurdles Remain

“We are very proud to say that we are now hardware agnostic. Our software can work on QWR devices, Oculus Quest, and other VR platforms. But as a medical device company, we need a hardware partner who can support regulatory compliance and long-term servicing,” Ramesh noted. 

By transitioning to domestic hardware partners, the company has managed to cut costs by nearly a third, making the solution more affordable.

Moreover, parental scepticism about VR-based therapy posed an initial barrier. Many parents questioned whether extended VR exposure was safe for children. 

“We explain that this is a temporary, structured treatment, not long-term VR exposure. It’s like a prescription, a controlled use of VR to improve vision function,” he clarified.

Looking ahead, NeuraSim aims to expand its VR therapy offering beyond amblyopia. The startup is developing AI-based retinal imaging tools and neurorehabilitation programs for conditions such as glaucoma and Parkinson’s disease. The company’s roadmap also includes expanding its clinical partnerships, increasing accessibility through home-based therapy solutions, and entering international markets.

“Our target is to be in 150-200 clinics by December 2025. Each clinic treating at least five patients per month for home therapy will allow us to reach a scale where vision therapy becomes more accessible,” Ramesh outlined.

With the rising number of patients, AI, telehealth, and VR-based therapy can help bridge the gap in healthcare access, ensuring more people receive timely treatment despite the shortage of specialists – a vision NeuraSim is actively working towards.

AI is everywhere, even in the database layer of a tech stack. However, given the sensitive nature of the sector, the application of AI in healthcare remains largely unexplored. Hyderabad-based MetaBrix Labs, an AI-powered 3D character generation platform, is working towards expanding its use case with healthcare providers. 

The startup aims to automate patient interactions by replacing boring chatbots and FAQs with a personalised 3D AI bot. In an exclusive interaction with Pratik Padamwar, the founder and CEO of MetaBrix Labs, AIM learnt about the idea behind the platform and much more.

The Initial Idea Behind Starting It All

To kickstart our conversation, we asked Padamwar about the inspiration behind creating the platform. He recounted that in 2022, his experience as a researcher in the generative AI space with companies like Intel, BMW, and more gave Padamwar the push to build a platform where people could create 3D models and characters and generate assets. 

It helped that the metaverse as a concept was in trend back then. But soon, he realised that it was not easy to do everything at once. Hence, he decided to focus on building something that could create hyperrealistic 3D characters from text or images.

What Are They Building Differently?

Many platforms are currently trying to generate personalised AI avatars. So, what’s different at MetaBrix Labs?

The answer lies in MetaBrix’s UltronAI, which claims to be India’s first platform that combines text, voice, and video to create highly personalised AI. As per the company’s statement, UltronAI’s proprietary technology enables rapid and cost-effective deployment of human-like digital personas, transforming engagement strategies across sectors.

“You have unicorns in the US, companies like Character.AI, Soul Machines, or Synthesia. They are all into video AI avatars. There is no real interaction happening,” Padamwar said. He further said that the interactive AI avatars they build try to hold conversations like humanised bots. 

When asked if he was inspired by Avenger’s Ultron character for his AI platform, Padamwar said that he was indeed fascinated by Tony Stark’s character to build something using AI.

Integrating 3D AI Bots With Healthcare: Privacy and Use-Case

When it comes to healthcare services, privacy is paramount. Padamwar said that while every healthcare provider has an HMS (healthcare management system), EHR (electronic health record), and CRM (customer relationship management), every component has its vendor. To integrate their 3D AI bot, they just need an API to extract data from patient records and provide more responses.

Most importantly, MetaBrix Labs does not store any data; it just acts as a visual representation on top of the data provided to its AI character. To implement this in a healthcare environment, all they need is a kiosk with a touch or non-touch screen, a single mic, and a dedicated 100 Mbps internet connection speed.

They have started deploying this in the Belenus Champion Hospital, Bangalore, and Apollo Hospital networks (in some of them). Through its implementation, the hospitals aim to reduce administrative tasks like insurance claims, automate interactions, free up time for providers, and give patients faster access to information. 

“By handling over 1,000 daily patient interactions and 45,000 conversations to date, the system has reduced administrative workload by 40%,” the company claimed.

It added, “UltronAI achieved a 90% patient satisfaction rating while managing tasks ranging from appointment scheduling to complex insurance queries in multiple languages, 24×7. The results at Belenus Champion Hospital are compelling: a 30% reduction in patient wait times and a significantly improved patient experience through AI-driven interactions.”

Padamwar also mentioned that in the immediate future, they aim to work with 500+ hospitals in tier 1, 2, and 3 cities to help providers automate patient interactions with companions that give real-time feedback. It is interesting to see AI’s integration in healthcare in ways that make a difference, no matter how small. 

India’s venture capital (VC) market is showing signs of recovery, but it’s still struggling with a significant gap in its AI ecosystem. Despite the rebound in overall VC funding – up to $11.2 billion in 2024 from a sharp decline in 2023 – India’s AI startups continue to face challenges. 

According to a Blume Ventures report, the VC market peaked at $37.4 billion in 2021 and declined sharply to $10.6 billion in 2023. The report further stated that India is the third-largest country in terms of the total number of unicorns, after the US and China.

While India is home to 117 unicorns, Krutrim AI remains the only AI startup that has reached unicorn status.

VC Sentiment for AI Startups 

Before the DeepSeek era, India was touted as the  ‘AI use case capital’. However, with the government’s strong push in this regard, the conversation has dramatically shifted towards building a foundational model. The IndiaAI Mission has reportedly already received 67 proposals, 20 of which plan to build LLMs.

According to AIM Research data, Indian AI startups raised a total funding of $560 million across 25 rounds in 2024, a decline of 49.4% compared to the previous year. For global AI startups, funding was around $27 billion from April to June.

AIM has extensively covered the significant investment opportunities available for AI startups in India.

Speaking with AIM last year, Prayank Swaroop, a partner at Accel, said that the 27 AI startups his firm has invested in in the past couple of years would be worth at least five to ten billion dollars in the future, which also includes wrapper-based startups. 

“A majority of people can start with a wrapper and then, over a period of time, build the complexity of having their own model. You don’t need to do it on day one,” Swaroop added.

Moreover, Peak XV Partners alone has set aside ₹16,000 crore, adding to the $20 billion available in VC funding. The VC ecosystem has $20 billion ready to invest in Indian startups. AI is currently the most significant theme, with investors showing strong enthusiasm for startups in this field. 

Another VC firm, Antler, has also invested in early-stage AI ventures with a $10 million fund. 

Meanwhile, incubators from Google, NVIDIA, and JioGenNext, along with government-backed GPU procurements, are also providing startups with resources to grow.

Indian-American venture capitalist Vinod Khosla, who is also an investor in Sarvam AI and OpenAI, also stressed the need for developing regional AI capabilities for national security, citing potential risks of foreign AI being restricted by sanctions.

In terms of resources, India has a strong AI talent pool, though some top professionals are in the US. In 2024, India ranked second in GitHub contributions to generative AI projects, just behind the US.

AI startups span the AI stack, from software applications to cloud infrastructure and hardware. Notably, AIM has an extensive list of GenAI startups.

What Is the State of the Global VC Market?

Unicorn creation peaked in 2021, with the US adding 305, China 44, and India 42. After 2022, numbers declined sharply due to a global funding slowdown. In 2024, India added six unicorns, up from two in 2023, but still lower than previous years. The US led with 56 new unicorns, while China added four.

Source: Blume Report 

US VC investments peaked at $363 billion in 2021 before falling to $152 billion in 2023 due to economic tightening. However, they bounced back to $191 billion in 2024, driven primarily by the AI, biotech, and fintech sectors. Large venture rounds (over $100 million) followed a similar pattern, recovering to $120.1 billion in 2024. AI alone accounted for $97 billion (or 80% of these deals during this period. 

China’s startup ecosystem, on the other hand, has faced challenges since its $85 billion peak in 2021, with investments dropping to $38 billion in 2024. This decline is attributed to regulatory crackdowns, geopolitical tensions, and economic slowdown. Similarly, the UK saw a significant drop in funding, from $38 billion in 2021 to $16 billion in 2024, due to economic uncertainty and declining investor confidence.

Real estate services company NoBroker unveiled ConvoZen.AI, a conversational intelligence platform that aims to transform customer engagement for businesses in various sectors. The move fortifies NoBroker’s entry into the Software-as-a-Service (SaaS) market with its AI-driven automation workflow. 

ConvoZen.AI is engineered to analyse and transcribe customer engagements across multiple channels, including calls, meetings, chats, and social media. 

The platform supports multiple languages, including English, Hindi, Tamil, Telugu, Kannada, and Marathi. By employing ML models, ConvoZen.AI offers features such as speaker identification, sentiment analysis, and entity recognition, enabling businesses to enhance customer service and operational efficiency.

“Unlike conventional solutions built using third-party models and optimised for Western accents, we built ConvoZen.AI to empower businesses operating at a Bharat scale to embrace Gen AI,” said Akhil Gupta, co-founder and CPTO of NoBroker. 

Google Cloud for ConvoZen.AI

The company has partnered with Google Cloud to use its Cloud AI infrastructure for the development and deployment of custom models optimised for large-scale operations. These models are trained on extensive datasets derived from NoBroker’s customer interactions, which include 45,000+ hours of contact centre conversations. 

Since its launch, ConvoZen.AI has been adopted by numerous companies across sectors such as lending, insurance, edtech, and e-commerce. The platform processes substantial volumes of customer interactions daily, providing businesses with insights that drive improvements in customer engagement and operational processes. 

Naren Kachroo, head of GTM, Google Cloud India AI, emphasised the growing importance of AI-driven agents during his keynote speech at the ConvoZen.AI Summit in Bengaluru. “2025 is going to be the year of AI agents. It is going to be the year of agenting AI.” 

He further elaborated on the significance of NoBroker’s entry into this space. “The work that NoBroker is doing with ConvoZen.AI is so important and relevant today because it ties to a business objective in a business workflow and accomplishes a task.”

Notably, ConvoZen.AI has significantly automated quality audits, leading to enhanced agent efficiency and more effective compliance tracking. Fifteen companies, including Cars24, LendingKart, LeapScholar, and Tata AIG, leverage ConvoZen.AI in their enterprise workflows.

NoBroker in SaaS Landscape

NoBroker’s entry into the SaaS market with ConvoZen.AI positions it among a growing number of companies offering AI-driven customer engagement solutions. The platform’s focus on multilingual capabilities tailored to the Indian market, combined with the credibility of its established real estate platform, provides a distinct competitive advantage. 

“It is wonderful to see ConvoZen.AI utilising the power of our Generative AI models toward creating new-generation ‘Agentic’ capabilities, both internally and for their customers,” said Manish Gupta, Director of Google Research India. 

Manish Gupta, Director, Google Research India at ConvoZen.AI Summit. Source: AIM

Interestingly, Wipro Limited has also partnered with Google Cloud to launch the Google Gemini Experience Zone. 

This initiative provides enterprises hands-on access to Google’s AI technologies, including Gemini models and Vertex AI. The Experience Zone enables businesses to experiment with generative AI applications such as natural language processing, image generation, customer interaction tools, and predictive analytics, facilitating the co-creation of tailored AI solutions to address specific industry challenges.

NoBroker aims to establish ConvoZen.AI as a leading conversational intelligence platform for mid-to-large enterprises. By making conversational AI more accessible and effective, NoBroker’s strategic shift into the SaaS domain with ConvoZen.AI reflects its commitment to innovation and addressing the dynamic needs of the market.

AI-powered wearables, in the form of rings, pendants or pins, have become the latest tech trend. Take the Humane AI Pin or the once-celebrated ‘Friend’ pendant, which once created a buzz but has now fizzled out. These innovations largely stem from the West, where AI hardware development is a well-established norm. In India, while the enthusiasm for AI wearables remains somewhat subdued, there’s a budding interest that hints at future growth.

Bengaluru-based AI startup NeoSapien recently received funding from angel investor Namita Thapar from Shark Tank India. The startup, founded by cousins Dhananjay Yadav and Aryan Yadav, is building AI pendants that they call the ‘second brain’. These devices are designed to capture and process information from conversations and allow users to perform tasks like booking cabs, drafting emails and much more. 

In an exclusive interview with AIM, Yadav said, “75% of the world’s data has been created in the last five years, yet our memory retention rate is only 2%.”

By capturing key moments, enhancing both professional and personal interactions, and even analysing emotional cues, NeoSapiens aims to help users regain self-awareness and make AI an intuitive part of everyday life. 

Why Isn’t Hardware Built in India? 

While the vision for NeoSapien is clear, the journey is far from straightforward. Dhananjay highlighted several challenges involved in building it, particularly when it comes to hardware development and adoption in India.

“Hardware is difficult because you need to be like a maniac,” he said, citing numerous hurdles that can arise, including hardware failures, software, and connectivity issues. 

Highlighting investors’ scepticism, Dhananjay recalled a recent conversation where an investor bluntly told him they were doubtful because no innovative hardware startup had ever emerged from India. 

Moreover, he noted that many Indian investors hesitate to back hardware companies due to a lack of confidence, as hardware is not built in India at the same scale as in other markets.

Moreover, Chinese brands are rapidly challenging Indian players in wearables, especially in audio devices and premium smartwatches. The brands are utilising their strong distribution across online and offline segments. 

Meanwhile, Dhananjay also pointed out that hardware requires a significant upfront investment, making it a high-risk venture. “Any deep tech, consumer tech, or hardware company requires investment. That’s where you need working capital,” he explained. 

He believes that addressing investor scepticism around Indian hardware startups requires government support, better funding mechanisms, and a shift in mindset. 

Dhananjay also emphasised the need for tax incentives and funding initiatives that specifically support Indian hardware innovation, citing examples from industries like space tech and EVs. “Unless and until we don’t get it out of India, we are never going to be a deep tech or hardware-first country,” he said. 

Another key solution, he believes, is creating Kickstarter-like platforms in India where consumers can support and pre-order innovative hardware products. “People in the US back products on Kickstarter. They are willing to test them out. But in India, we lack such platforms.” According to him, this cultural gap makes it harder for Indian hardware startups to gain early traction. 

With hardware posing numerous challenges, building AI wearables is only going to get more difficult. The Humane AI Pin, built by former Apple executives, serves as a reminder that nothing is immune to failure. Despite these challenges, some Indian startups have ventured into building AI wearables. 

Dhananjay remains optimistic that successful Indian hardware startups will gradually change investor perception, proving that original innovation can thrive in the country.

West Continues Building 

Big tech companies continue to remain bullish in the wearable segment market. OpenAI recently filed a patent to develop smartwatches, jewellery and wearable cameras. 

In a recent episode of All-In Podcast, investor and entrepreneur Naval Ravikant, known for his early-stage investments in Uber, Twitter, and other notable companies, said, “Now, I’m building a new product, and this time, I’m going into hardware. I’m just building something that I really want.”

Ravikant reveals that his interest stemmed from Elon Musk’s work. “The guy can be the Diablo player and do Doge and run SpaceX and Tesla and [The] Boring [Company] and Neuralink; I mean, it’s incredibly impressive,” he said.  

Despite the challenges, startups like NeoSapien, inspired by global visionaries like Musk, are paving the way for AI wearables in India. With changing mindsets and increased support, India’s hardware innovation potential is on a positive track.

In a world where AI innovation is largely dictated by Silicon Valley and China, Camb.ai is quietly making a case for deep-tech leadership from outside the usual power centres. Founded by Avneesh Prakash (CEO) and son Akshat Prakash (CTO), the company operates from Dubai but has a significant Indian influence. 

Camb provides speech and translation AI offering, dubbing content in over 140 languages using its proprietary AI models. The startup has partnered with sports, production companies, and other verticals for live AI content. 

Rethinking AI Economics 

While most AI companies are locked in a race to scale large models, Camb is betting on small language models (SLMs). Its focus on SLMs through models such as Mars 6, which has just 80 million parameters, makes it apt to run on a smartphone. 

“If I can fit something on your phone, it suddenly changes the game completely. You don’t need these large-scale GPUs. You don’t need to have a central place to call everything from,” said Akshat in an exclusive interaction with AIM. 

Akshat believes that this approach reshapes AI’s cost dynamics. Instead of companies paying per-API-call fees to OpenAI or Google, Camb.ai envisions a future where AI models run independently on consumer devices. 

“We are so much more than speech-to-speech or AI dubbing. When we are successful, we also have the ability to redefine the economics of AI,” he said.

AI Without Borders

Unlike western AI firms that operate from a singular tech hub, Camb.ai has built its foundation with a globally distributed team. “People are very confused about how a Dubai-born or an East-born company could be pulling off such feats that they would have expected western companies to do,” said Akshat. 

Beyond its global presence, Camb.ai also challenges the narrative of AI ownership. While many companies rely on external APIs from OpenAI or Anthropic, Camb has built its own proprietary models. 

As a strategic move to prove its independence, it even open-sourced Mars 5, a high-performance voice model. “The world didn’t want to believe in us, so open-sourcing was a great way to tell them, ‘Look, we’re building our own technology’.”

Camb.ai partnered with Major League Soccer and became the first company to live stream the games in multiple languages. The company’s real-time AI dubbing and translation were also successfully deployed during the 2025 Australian Open. 

“Sports has all the difficult elements of content that make this problem hard. It has background cheering, noise, loud environments, multiple commentators, niche lingo, and fast-paced action,” Akshat explained. 

By perfecting AI dubbing for sports, Camb ensures its technology is robust enough for broader enterprise applications.

That’s not all. The startup, which raised a $4 million seed round led by Courtside Ventures last year, has also partnered with IMAX to offer multilingual support for films. 

Beyond Dubbing

While Camb started with AI dubbing, its long-term vision is to break language barriers at scale. “The internet was made for English speakers, and we want to redesign it for the world,” said Akshat. Its AI solutions now extend beyond speech and video translation, helping enterprises localise digital content, improve fan engagement, and automate multilingual communications.

The company’s app, Savant, takes this mission further. It allows users to type and interact in multiple languages seamlessly, making cross-language communication frictionless. “If anybody can text you in the language they’re most comfortable in, and you can receive it in your preferred language, that’s amazing,” he added.

Despite its rapid rise, Camb.ai is still competing against AI giants with massive funding and infrastructure. However, Akshat believes their focus on smaller, more efficient AI models will give them an edge in the long run. 

“The biggest challenge in AI today isn’t just innovation, it’s making AI sustainable and scalable. If we can solve that, we change the game,” he concluded. 

The dawn of AI agents promises a shift in how enterprises build and deploy applications. By now, we’ve seen big tech and SaaS giants going all-in on agents. However, other companies have been quietly working on providing AI solutions for enterprises and have also rolled out no-code AI agents. Case in point: Kore.ai. 

The eleven-year-old Orlando-based Kore.ai, founded by Raj Koneru, has its second major hub in Hyderabad, which includes an R&D centre. The company helps businesses create AI chatbots and virtual assistants, and its customers span banking, healthcare, and airlines. 

Kore.ai is now enabling companies to build and deploy AI agents without extensive coding knowledge. 

Conversational AI to AI Agents 

“We provide a platform which is like a set of platform core capabilities, services, and a lot of no-code tools for builders as well,” said Prasanna Arikala, CTO of Kore.ai, in an exclusive interaction with AIM. 

The platform has evolved from conversational AI to a comprehensive agent development ecosystem. It now includes the ability to build agents, create tools for agents to interact with, and develop sophisticated RAG (retrieval-augmented generation) pipelines for multi-agent applications.

“If you have a system of record, you can build pretty much any application with agents. The application development and deployment paradigm is significantly changing, and enterprises have quickly realised that,” explained Arikala, emphasising the paradigm operational shift of enterprises. 

On the database front, in an earlier interaction with AIM, Redis VP of AI product management Manvinder Singh confirmed that Kore.ai uses Redis as a data platform to power their virtual AI agents. 

Kore.ai’s AI agents are making significant impacts across various industries. The company boasts about 450 customers, including some of the world’s largest banks and healthcare companies. 

One of its customers, a major wealth management company, deploys AI agents for its 60,000 employees. Arikala explained how these agents have dramatically reduced the time required for tasks such as creating customer proposals. 

“Previously, it would take weeks for them to compile, collect all the information in accordance with the enterprise guidelines, corroborate, build, templatise, and deliver a report. Now, they just upload the data, and it gives the outcome within minutes,” he said. 

Arikala claims that the platform’s versatility allows it to be applied across various domains, including customer service, process automation, and enterprise work management. 

“One out of the top five banks in the US deploy AI agents for their customer service,” Arikala said. 

During the COVID-19 pandemic, Pfizer leveraged Kore.ai’s AI agent platform to deploy a multilingual support system in 17 languages globally. This system assisted healthcare professionals in efficiently accessing critical vaccination-related information.

Challenges Remain

Despite the platform’s success, Kore.ai acknowledges the challenges of deploying AI agents at scale, particularly the governance part when the number of agents keeps growing. 

Arikala emphasised the need for oversight in agent development, questioning who built them, how they are being used, and whether they comply with enterprise guidelines and SOPs. Unlike workflows, agents don’t follow a deterministic approach, making safeguards essential.

To address these challenges, Kore.ai is developing solutions, such as a built-in agent evaluation service, as part of its platform. It allows for periodic assessments of AI agents, generating comprehensive reports on their performance and behaviour. 

Kore.ai envisions a future where AI agents become ubiquitous in enterprises. “In the future, the enterprise will be all about a network of AI agents, and there will be centralised orchestrators that allow for a hub for the internet sorts, “ predicts Arikala. 

As confirmed by Arikala, Kore.ai has been witnessing growth rates of 100% year-over-year for the past three years and continued significant growth for the current fiscal year. With this growth rate and IPO plans in the coming years,  it seems that Kore.ai is well-positioned to lead the AI agent revolution. 

We have all heard the saying – cricket is not just a sport; it’s an emotion. With a global audience of 2.5 billion, including 600 million viewers in India alone, cricket stands as one of the most-watched and widely played sports. While outdoor cricket remains the ideal choice, the passion for the game has seamlessly transitioned indoors with interactive and immersive experiences shaping the gaming format.

Bengaluru-based startup MetaShot is redefining home entertainment by merging cricket with immersive gaming technology. Co-founded by Prince Thomas, Ranjit Behera, and Ajith Sunny, Metashot has developed an innovative hardware-based gaming system that allows users to play cricket in their living rooms. The idea stemmed from their love for the sport and the realisation that motion-based gaming remained largely expensive and inaccessible in India.

Launched in September 2023, the startup claims to have sold out two months’ worth of inventory in just 10 days. So far, MetaShot has sold approximately 25,000 units, with revenue growing fivefold year-over-year. 

Unlike traditional gaming consoles that rely on expensive hardware, MetaShot’s system integrates with a smart bat and a motion-tracking sensor. “Whatever you do in your living room, your avatar will do the same thing on the screen,” Thomas said, comparing the system with a Nintendo Wii-like experience at a fraction of the cost. 

This innovation has also struck a chord with parents. “We’ve had many testimonials from mothers who are happy that their kid is at least having some kind of physical exercise,” Thomas added.

Gaming with Smart Tech

MetaShot’s cricket simulation works without cameras or infrared tracking, unlike traditional gaming consoles such as Xbox Kinect or VR headsets, which use multiple cameras for motion detection. Instead, it relies solely on 9-axis sensors embedded in the smart bat, which track a player’s movements, shot angles, and power. 

The data is processed using a smartphone, tablet, or laptop, eliminating the need for an expensive console.

Interestingly, MetaShot has expanded its customer base, with 37% of users being both children and adults, 33% being adults alone, and 30% kids. 

“There are adults playing, they use it for party games, they connect with their friends, they buy in bulk, three or four friends will buy and they will play together. So, we are increasingly seeing those kinds of use cases,” said Thomas.

Made in India 

One of MetaShot’s core principles is to manufacture locally. “From the beginning, we were very clear [that] we would try to make it in India,” Thomas said. While sensors are imported, the bat moulds, components, and assembly are done entirely in Bengaluru. This decision aligns with India’s push for domestic manufacturing and avoids reliance on Chinese components, despite the challenges of building a hardware ecosystem locally.

Metashot has also expanded its retail presence, partnering with Blinkit for quick deliveries, making their product accessible within minutes for last-minute purchases, especially for parties and group play. 

Long Way Ahead

As of January 2025, MetaShot has raised approximately $2.19 million across multiple funding rounds, including a ₹11 crore seed round led by Sauce.vc. Moreover, the startup secured ₹1.6 crore for a 5% equity stake in Shark Tank India.

Despite its unique positioning, MetaShot competes with major global gaming brands. Products like the Nintendo Wii and Xbox Kinect introduced motion-based gaming years ago but failed to gain traction in India due to high costs. 

“A similar product at this price point doesn’t exist,” Thomas pointed out. While a Nintendo Switch or PlayStation costs between ₹40,000 and ₹50,000, MetaShot’s cricket game is priced significantly lower at ₹5,000, making it far more accessible.

Gaming giants such as Sony, Microsoft, and Nintendo dominate the high-end console market, selling 2-3 lakh units annually in India. Thomas believes that if MetaShot reaches its full potential at a lower price point, the company could surpass those numbers. “If we start hitting our potential, we can do much more.”

MetaShot is not alone in the sports-tech gaming market. Companies such as StanceBeam and Freebowler have developed smart cricket equipment, while Motion Sense AI is working on AI-powered motion tracking for gaming. 

As MetaShot expands, its next frontier includes multisport gaming. The team is already developing a universal motion-tracking device that allows users to switch between cricket, tennis, and other sports by simply changing the bat or racket attachment. 

“We are working on a universal device that lets you play multiple games. Most likely, we will launch it [in the] next financial year,” Thomas concluded. 

Sridhar Vembu, former CEO of Zoho Corp, recently highlighted the need to stop glorifying English in India’s R&D ecosystem. “There is a lot of R&D talent in India if we get rid of the English barrier and the social stigma of not knowing English well,” he posted on X. 

“I am right now working with extremely capable engineers on some advanced tech, and we converse in Tamil because that is what they are most comfortable with,” he added.

While Vembu’s emphasis on local languages was one aspect of the discussion, the localisation of AI solutions is gaining momentum. Indian companies are increasingly preferring homegrown AI startups over big-tech firms.

The Preferred Choice 

In a past interaction with AIM, MN Anucheth, the JCP of Bengaluru Traffic, spoke about the traffic department working with several homegrown AI startups to leverage AI solutions. 

“Since Bengaluru is the tech capital of India and a lot of AI-based startups are based in the city, we have been lucky enough to be able to work with many such companies,” he said. “AI has been made accessible to us, for which we would otherwise rely on some foreign import or off-the-shelf product, which generally do not work in real-time conditions.”

The Bengaluru Traffic Police has collaborated with many Indian startups to enhance AI-driven traffic management. For instance, Monday Technologies supports AI avatars for awareness videos and drone-based monitoring to detect road blockages and accidents. Other key partners include IBI (the developer of ASTraM), Skita, and Matrix Technologies, with Videonetics as the OEM. 

Anucheth explained that continuous feedback helps refine models, such as improving seatbelt detection accuracy. Though big-tech firms such as Google and Cisco offer traffic management solutions, authorities prefer to maintain flexibility and control over their infrastructure.

“Nothing against big tech companies, but I think our experience has been that we can’t work with them to give tailor-made solutions to us,” said Anucheth. 

Nuanced Approach

From an investor’s perspective, VCs look to invest in Indian AI startups that offer nuanced solutions that big-tech companies cannot deliver. 

Citing Google’s PaLM models as an example, Capria Ventures’ explained to AIM how startups have an edge in understanding vertical-specific needs. In the health sector, especially hospitals, where the radiology department requires in-depth analysis, a local player has a better edge.  

“The big-tech models are going to be there to prove the science of the underlying technology they have, but they are not good at solving the vertical needs of what the radiology department at hospitals need, top to bottom,” said Will Poole, co-founder and managing partner of Capria Ventures, in an interview with AIM earlier.

Poole cites 5C Network, which has spent seven years developing a specialised, end-to-end solution tailored for radiology departments. While AI models play a role, they are just one component of a much larger system.

Big tech may develop powerful models, but access to high-quality, diverse medical imagery is essential for training effective AI. “What Kalyan, the founder of 5C, built was a network of hospitals from which he takes images. A radiologist can read those images… He’s done 11 million of them, adding, I don’t know, half a million per month,” Poole said. 

“Without that kind of data… you’re not going to be able to build an AI that’s as good as somebody who has it,” he added. 

Poole explains that for investors, the key lies in backing founders with proprietary data access, enabling them to scale rapidly and outpace competition. 

Indic Models

Navana AI, a Bengaluru-based voice AI startup that develops indigenous AI-powered speech recognition and NLP solutions, told AIM that big tech companies posed a lot of problems when it was in the research phase of trying out interfaces with NLP. 

“We started using Google, Microsoft, AWS, all of the existing services that were out there, but none of them worked for Indian languages at that time. Even today, most don’t work for non-major languages or low-resource languages,” said Raoul Nanavati, co-founder and CEO of Navana ai. 

While big-tech companies have been collecting and building language data, it has mostly been English and Western languages such as French, German, and Spanish, as these native language users have been on the internet for decades. 

This has been a challenge for these companies when they build for India. “So there’s a cold start problem in India for language AI,” said Nanavati. 

Recognising this, the team decided in 2018-19 to develop indigenous speech recognition AI and NLP solutions for all Indian languages, overcoming the data scarcity that had hindered progress in this space.

With the nuanced approaches each startup offers, enterprises are actively seeking to collaborate with Indian AI startups, a trend that shows promising potential.

By now, you must know that China’s latest AI model, DeepSeek-R1, has been the centre of all conversations for having built a SOTA model with scant resources with respect to compute and cost. It shattered the idea that building a model as capable as OpenAI’s on a $10 million budget is impossible—remember Sam Altman’s last visit to India? 

With DeepSeek setting a precedence for everyone, India has gotten the boost it always needed. 

India’s DeepSeek Ambitions

When US President Donald Trump announced Project Stargate a week ago, India got talking about building in the country. India building its own Project Stargate was portrayed as a necessity, with many tech leaders weighing in on the conversation. 

The discussions had barely died down when DeepSeek brought forth the next idea of why India can’t build one just like it.

Well, India is already building it. 

“Yes, we definitely are! It won’t be a 671B parameter one (to begin with), but it’ll be a frontier model in its parameter category,” said Abhishek Upperwal, founder and CEO of Soket AI Labs, in an exclusive interaction with AIM.  

“The pace of development will depend on the kind of funds we get access to, but we are gonna definitely build it,” the founder of the Gurgaon-based AI research startup added. 

Upperwal stated that Pragna-1B (Soket’s AI model) marks the team’s initial step toward developing frontier models. The 1.25 billion-parameter model was trained on a budget of just $100K, covering both synthetic data and compute costs. 

“The plan is to bootstrap bigger models using smaller ones and any open-source model with a permissive license—while keeping compute costs dirt cheap,” he said. 

He highlighted that high-quality data and training optimisations make this approach feasible, pointing to DeepSeek as a successful example.

Upperwal noted that if  “less resources” translates to $2-3 million, the prospects for building frontier models are either bleak or significantly slow. In such a scenario, companies would have to prioritise revenue-generating products over AI model development.

“I think we need at least $10 million to start working on frontier tech, and this money should be purely dedicated to R&D for building these models—no distractions like building applications or even thinking about GTM. This is where investors and founders need to align with patient capital,” said Upperwal. 

Similarly, Reliance-backed Indian AI startup TWO AI is building a cost-efficient multilingual AI model family with speech, search, and visual processing in 50+ languages. It believes it has already been building DeepSeek-like models. 

“DeepSeek’s RL-only post-training approach and insights like distilling reasoning into smaller models really resonate with what we’re doing at TWO AI,” said Pranav Mistry, founder and CEO of TWO to AIM. 

Mistry believes the AI race now demands rapid innovation rather than massive compute power. “Gone are the days when you needed a 20,000 GPU farm to train a single model,” he said. 

He added that TWO AI has demonstrated this with its SUTRA model, which outperforms SOTA models in the official MMLU for Indian languages despite being trained on a $2 million budget. 

While greater resources can accelerate innovation, optimised approaches are proving just as crucial. “Of course, more resources can help accelerate the speed at which we can innovate,” he added. 

Pratyush Kumar, co-founder of Sarvam AI, another Indian AI startup that is developing LLMs and GenAI solutions for Indian languages, recently posted on X inviting Perplexity co-founder Aravind Srinivas to join their mission. 

“Aravind, at SarvamAI we are building sovereign models that combine deep reasoning and Indic language skills. Would love to have you join this mission!” he wrote. However, when AIM reached out, Sarvam AI declined to comment on DeepSeek. 

Multimodal AI platform Krutrim AI, started by Ola’s Bhavish Aggarwal, is also on a mission to cater to the Indian audience via their multilingual platforms. 

What is Stopping India? 

“Very soon, we will also have our own LLMs,” said IT minister Ashwini Vaishnaw, at the recent Utkarsh Odisha Conclave. “In the India AI compute facility, we have received compute bids for creating 18,000 GPUs,” he said. 

While the government is slowly encouraging and providing incentives to promote AI in India, VCs are still sceptical about investing fully in it. 

“The problem is that the benefit here isn’t immediate revenue generation, which is why VCs run away from these kinds of ventures. But the real ROI is in gaining the know-how of building intelligence at scale, which can create value in a hundred other ways (just imagine the kind of leverage DeepSeek holds today),” said Upperwal. 

“Intelligence and the know-how to build one will be the most valuable IP in the future,” he added.

Upperwal believes that to reach DeepSeek R1’s level, we will need at least $50 million. “DeepSeek is already on its 3rd version, plus multiple other models. The cost to get here should be the aggregate of everything they’ve spent so far. I’d estimate $50-100 million,” he said. 

He believes the key lies in securing adequate R&D funding (ranging from $5-10 million per startup) for at least 4-7 teams. “Sarvam is the only startup with access to such funds, but it’s splitting its focus between figuring out use cases and building models, which slows down progress,” he said. 

In a blog post, Zerodha co-founder Kailash Nadh shared his views on DeepSeek, focussing on research and human capabilities as a priority. 

Nadh believes that India’s AI sovereignty and future depends not on a narrow focus on LLMs or GPUs but on building a foundational ecosystem that encourages breakthroughs through a blend of scientific, social, and engineering expertise across academia, industry, and civil society. 

“In fact, the bulk of any long-term AI sovereignty strategy must be a holistic education and research strategy. Without the overall quality and standard of higher education and research being upped significantly, it is going to be a perpetual game of second-guessing and catch-up,” he said. 

OpenAI finally introduced the ChatGPT moment for AI agents with its new AI agent ‘Operator’. The agent can perform tasks on the web without human intervention based on users’ instructions. Notably, the focus of every enterprise and startup is on AI agents capable of performing tasks independently. Indian founders are no exception as they advance into the next phase of AI agents, now powered by voice capabilities.

The Rise of AI Voice Agents

Moving from text-based interaction to using voice to activate tasks and agents to run them is a trend that AI startups in India are actively pursuing. 

Sudarshan Kamath, founder of smallest ai, which builds text-to-speech models and voice agents, shared his views on voice agents. The company’s journey into voice AI began with the realisation that everyone has a very different voice that they like. To address this, smallest.ai introduced voice cloning, which allows users to create customised voices with reference audio. 

Smallest.ai’s focus on AI agents is rooted in their potential to handle complex tasks in real time. “There are companies who are moving away from IVR-based systems to voice bot-based systems, and these voice bot-based systems are smarter, more interactive, and more realistic,” Kamath said while interacting with AIM. 

He explained the use case of these voice agents in content creation, such as companies producing product videos or marketing campaigns. “Or, it could be individuals who are influencers or social media accounts who are basically trying to create content on Instagram, YouTube,” he added. 

Why Voice? 

Kamath highlighted how large enterprises, including publicly listed ones, are increasingly exploring voice-based workflows. “This shift has happened because generative AI has made these voices much more realistic while maintaining very low latencies,” he said. 

Kamath also believes that voice-based solutions offer a high return on investment (ROI) and significantly enhance engagement and user experience. “So, investors are fairly bullish about the voice as the market itself is going to grow.”

Bengaluru-based conversational AI and voice automation startup Gnani.ai claims to currently handle 30,000 concurrent conversations and a few million voice AI conversations daily. Their voice-first SLMs for Indian enterprises are trained on millions of audio hours and billions of Indic language conversations. 

“Indian AI startups are focusing on building voice agents due to the country’s diverse linguistic landscape, the rapid adoption of smartphones, and the increasing demand for seamless customer interactions across industries,” Ganesh Gopalan, co-founder and CEO of Gnani.ai, told AIM. 

“The rise of vernacular voice interfaces also aligns with the push for digital inclusion in India, enabling startups to cater to a broader audience while tapping into the growing demand for localised, AI-driven solutions,” he added.  

Gnani.ai caters to industries such as banking, finance, and insurance and helps them use AI-powered solutions for tasks such as customer support, lead qualification, EMI collection, and insurance renewals.

“Some focus on multilingual support with high accuracy in regional languages, while others emphasise industry-specific solutions, such as BFSI, healthcare, or retail, tailoring their AI to address niche requirements,” he said.

Another Bengaluru-based voice AI startup, Navana.ai, develops indigenous AI-powered speech recognition and natural language processing (NLP) solutions. Having worked with institutions such as IISc Bangalore, IIT Madras, and Bhashini on open-source data collection efforts and co-authoring academic papers. Their voice agents are integrated into applications for industries such as BFSI, agriculture and government services. Ujjivan and Bajaj Finserv are a few of their customers.

“In the last year and a half to two years, the big shift came when LLMs came around and made telephony a very viable channel to reach all of India and plug in AI to do all sorts of services,” Raoul Nanavati, co-founder and CEO of Navana ai, said during an interaction with AIM. 

Nanavati believes that voice agents are gaining traction because they make digital services accessible to first-time internet users and address India’s linguistic diversity. “None of them [Google, Microsoft, AWS] worked for Indian languages at that time. Even today, most don’t work for non-major languages or low resource languages,” he emphasised. 

What Next? 

With voice gaining prominence as a key mode for AI agents, the focus is now probably shifting toward identifying the next trend in the field.

“After voice agents, the next trend in AI is likely to revolve around multimodal AI agents that integrate voice, text, and visual interactions for more immersive and context-aware experiences,” Gopalan said.  

He believes that such systems can improve user engagement and create a more intuitive and enriched interface.

“Additionally, the focus will shift toward hyper-personalisation powered by generative AI, where conversational agents predict and adapt to user needs in real time,” he concluded. 

Similarly, Praveer Kochhar, co-founder of Kogo Tech Labs, which recently unveiled universal voice assistants for automobiles, believes the agentic systems will move towards larger goal accomplishment. “This transition from task-oriented agentic flows to goal-oriented flows is the next big thing that you’ll start seeing, whether it’s in front office, back office, direct to customers, everywhere,” he told AIM.

Talking to your car isn’t a new concept anymore. Voice assistants like Google Assistant and Siri made this possible years ago. Now, an Indian AI agentic assistant is taking this innovation to the next level and aiming to transform the driving experience for millions with an agentic mesh framework.

Bengaluru-based AI startup Kogo Tech Labs recently unveiled India’s first universal voice assistant for automobiles at the Bharat Mobility Global Expo 2025. A few months ago, the startup announced an AI agent store offering AI tools, agents and plugins.

“Unlike Siri and Alexa, these agents not only give you access but can also operate apps and services on your behalf,” said Raj K Gopalakrishnan, co-founder and CEO of Kogo Tech Labs. 

Unlike existing voice assistants such as Siri or Alexa, Kogo’s assistant operates at a deeper level, directly interacting with applications to perform actions such as initiating navigation rather than opening a navigation app alone. It can execute a variety of commands, from controlling car hardware (e.g., switching on wipers or adjusting air conditioning) to managing apps and services.

“This assistant is not just a showcase but a practical application of what’s possible with our technology. Whether it’s a truck driver topping up a FASTag or a family planning a vacation, the assistant provides a versatile, multi-language platform to meet diverse needs,” Gopalakrishnan noted.

According to him, the new capability extends across domains, from navigation to booking airline tickets and provides users with a unified, intelligent assistant that can handle diverse tasks seamlessly.

Enter MapmyIndia 

As a voice for automobiles, navigational capabilities become critical and, to address that, Kogo has strategically partnered with geo-intelligence company MapmyIndia.

Through the partnership, MapmyIndia’s advanced geo-intelligence stack, along with Kogo’s AI assistant, offers navigation and location-based services. Kogo’s assistant will be able to support a wide range of automotive applications, from real-time navigation to enterprise logistics. The partnership will also provide access to MapmyIndia’s extensive customer base, including 30 original equipment manufacturers (OEMs) and numerous enterprise clients. This will position Kogo to scale its solutions effectively.

“Mobility, by definition, requires geo-intelligence, and who better than MapmyIndia? They are the leaders in geo-intelligence, at least in India. They also have a lot of deep learning and access, through their partnership with ISRO and NAVIC, etc,” Gopalakrishnan highlighted. 

Notably, a few years ago, MapmyIndia acquired more than 26% stake in Kogo Tech Labs. 

Google, ChatGPT? 

Big tech companies have partnered with automobile manufacturers to host their cloud and assistant features. Recently, Google Cloud announced its agentic integration with Mercedes. Similarly, ChatGPT is also integrated with Mercedes. 

Drawing parallels, Gopalakrishnan explained that while they are catering similarly, their platform works on an agentic mesh. “So right now, what this means is that it’s not dependent on one service. It potentially gives access to millions of apps and businesses because they can all now talk to each other. So, we’re platform agnostic,” he said.  

Kogo is already testing its platform with four OEMs – two in North America and two in India. “It’s a similar kind of approach where we will go in right at ground zero and implement this.”

They are also working on integrating their platform at a stack level with major semiconductor players, which will be announced in the coming months. 

Way Ahead

While voice assistants have become a popular trend, Kogo’s leaders caution against viewing them as a one-size-fits-all solution. “Voice agents are fascinating, but their true potential lies in specific use cases where they outperform traditional interfaces, such as while driving or during hands-free tasks,” said Praveer Kochhar, co-founder of Kogo Tech Labs. 

He predicted that by 2025, real-world deployments of voice technology at scale will become more commonplace.

Looking ahead, Kogo is focusing on enhancing the cognitive capabilities of its agents. “We’re working on rolling out goal-oriented agents in three months, capable of solving complex problems independently,” Praveer shared.

He also envisions a shift in AI capabilities from task-oriented to goal-oriented frameworks. 

“You will see a huge capability in the cognitive ability of agents and tools, which means they will be able to solve complex tasks. They will be able to chart out their own pathway of solving problems,” he concluded.  

Imagine watching a dubbed movie, but the characters’ lip movements stubbornly stick to the original language, completely breaking the illusion that movies are meant to be. While dubbing, translation, and subtitling may be sorted, if this aspect is ignored, the movie experience is dampened. 

In comes NeuralGarage, which is on a mission to transform the dubbing industry.

AI Facial Sync

The founders of NeuralGarage

Bengaluru-based NeuralGarage was founded in 2021 by Anjan Banerjee, Subhashish Saha, Subhabrata Debnath, and Mandar Natekar, who have known each other for decades. The platform is looking to solve the problem of mismatched facial movements in dubbed content. 

NeuralGarage was one of seven Indian AI startups selected for the AWS Global Generative AI Accelerator Program. It received $1 million in AWS promotional credits, among other support initiatives. 

Its technology, VisualDub, not only addresses lipsync movements but also ensures the final product looks natural and high-quality. “The challenge is maintaining the video quality when adjusting facial expressions, especially with no limitations on camera angles and lighting,” the company stated. 

“It cannot look AI-ish,” it emphasised, particularly when working with renowned actors whose facial movements are unique.

“For example, if you’re watching Money Heist, the professor is still speaking in Spanish while you are listening to him in English. What we do is change the facial movements of the actor so it looks like he has spoken in English,” explained Subhabrata Debnath, co-founder and CTO of NeuralGarage. 

Advertising and Movie Industry 

While dubbed content has become popular, especially on streaming platforms such as Netflix, NeuralGarage sees an opportunity to further enhance the viewing experience. “Once you see it with sync, it is very difficult to go back,” they noted. 

Prime Video’s global head recently acknowledged this challenge, stating that content is still not consumed well because facial expressions and lips don’t always match.

NeuralGarage is already being leveraged by major players such as Amazon India, Coca-Cola, and others in advertising, allowing brands to repurpose content with different messages. Debnath explains advertisements shot in December can be updated in February to reflect changing discounts or festivals. 

“Amazon India recently started using our solution to repurpose the same video footage with different messages for seasonal campaigns,” said Debnath. 

In the film industry, NeuralGarage collaborates with movie distributors, such as Europa Movies, which has a 95% market share in India. Through these partnerships, they address challenges related to global distribution, piracy prevention, and dubbing with realistic synchronisation.

Hurdles Ahead

NeuralGarage faces significant challenges in perfecting its dubbing and facial synchronisation technology. A primary issue is ensuring that facial adjustments seamlessly match the original high-quality video. The technology also needs to adapt to varying camera angles and lighting conditions, as films are not shot with AI’s constraints in mind.

Maintaining a natural, non-AI-generated look is especially crucial for high-profile actors. “Shah Rukh Khan’s face is his identity. For movie stars, the expectations are very high,” he said.

The scalability of the solution presents another challenge, with films often exceeding 300 GB in size. 

However, the challenge of competition persists. Additionally, several AI startups, both globally and in India, are working in the same space. Dubverse.ai and Vitra AI are a couple of names. However, NeuralGarage is looking to go big in 2025. 

The platform plans to boost its visibility in 2025 by actively investing in marketing and sales, transitioning from its reliance on organic word-of-mouth within the post-production industry. “This year, we will invest in marketing and sales so that the solution reaches more hands,” he said. 

After raising $1.45 million from Xfinity Ventures in 2022, the company is preparing for a Series A funding round to scale operations and refine its product. The company’s goal is to ensure that more businesses try its solutions, which streamline subtitling and localisation processes. 

Computer-aided design (CAD) is considered the backbone of the construction industry, but the global market doesn’t seem to believe this. While CAD is indispensable in India, the US has restrictions that prevent vendors from relying on it, causing construction inefficiencies, especially in landscaping. 

An Indian AI startup has come forward to address this problem on US soil. 

Why USA, Not India? 

Founded in 2017 in Delhi by three friends from IIT-D, Attentive.ai provides AI-powered solutions for the landscaping and outdoor services industry. 

Attentive.ai has deliberately focused on North America, citing distinct market dynamics as the reason for bypassing the Indian market for now. In the US, takeoffs (measuring site areas) from PDFs are standard due to intellectual property concerns, making the process more labour-intensive. 

In contrast, Indian firms often use CAD files, which allow for quicker and easier measurements.

“The Indian market operates differently. CAD files are more widely circulated here, which simplifies takeoffs. In the US, PDF files dominate, and extracting measurements from them is significantly harder,” explained Shiva Dhawan, co-founder and CEO of Attentive.ai, in an exclusive interaction with AIM. 

This fundamental difference in workflows makes the US a more suitable market for Attentive.ai’s solutions. 

Additionally, Dhawan pointed out the economic advantages of targeting North America. “We aren’t focusing on India in the short- or medium-term because the US market offers better opportunities and cost arbitrage,” he said.

Attentive.ai initially focused on AI services, building expertise in computer vision. However, a strategic pivot in 2021 led the company to develop SaaS solutions tailored for the commercial landscaping and construction industries in the US and Canada markets. 

“We realised services were a means to an end, and recurring revenue is where valuable businesses are built,” Dhawan said.

The Promising Landscaping Industry

Attentive.ai’s flagship product, AutoMeasure, automates takeoff processes for landscaping companies, while Beam serves construction firms. 

Traditionally, takeoffs required hours of manual effort with drawing tools. Professionals would click and trace polygons to calculate dimensions manually, which was tedious and time-consuming. Attentive.ai’s platform eliminates this hassle by using AI to automate 60% of the work. Human experts verify the remaining for quality assurance.

“The way it works is simple. Customers upload a site plan or input an address, and within hours or a day at most, they receive accurate measurements,” Dhawan explained. This process not only improves efficiency but also enhances accuracy, enabling firms to redirect resources to more critical tasks. 

“Our tagline is ‘get time back,’ as we save companies significant time, allowing them to focus on other activities,” he added.

This efficiency has driven rapid adoption in North America; Attentive.ai now serves over 1,000 clients in the US and Canada. The company’s ability to cater to an underserved yet essential market has set it apart in a competitive landscape.

Addressing a Niche Market

Its success stems from its ability to address a niche yet massive market. AI companies have historically underserved the billion-dollar landscaping and construction industries. Attentive.ai identified this gap and leveraged its expertise in computer vision to create tailored solutions.

“These industries have huge potential but are largely ignored by technology firms. If you can build for them, it creates a moat,” Dhawan explained. This focus on niche AI applications, rather than saturated areas such as call centres, has differentiated Attentive.ai and helped them attract significant investments.

Since its pivot to SaaS, the company has raised $18 million from PeakXV Surge, Vertex Ventures, Tenacity Ventures and others. “These investors believe in our vision of AI-enabled services transforming trillion-dollar industries,” Dhawan said.

Challenges Remain

While the company has made impressive strides, it faces its share of challenges. Chief among them is talent acquisition. “Finding good talent and convincing them to work in an unconventional industry is tough,” Dhawan admitted. 

The niche nature of the landscaping and construction sectors often makes it harder to attract top-tier talent. Despite this, Dhawan believes the company’s outsider perspective has been an advantage. 

“I’m not from the construction industry or a civil engineering background, but that allows us to see problems more clearly. Our team includes both industry outsiders and experts, which helps us balance fresh perspectives with domain expertise,” he said.

Another challenge is maintaining customer trust in an industry historically reliant on manual processes. However, Dhawan highlighted that the shift is already underway. “Many firms that previously managed takeoffs in-house now outsource them to us because they trust our service and see the time savings,” he noted.

Attentive.ai’s future plans are focused on scaling current offerings rather than diversifying into new products. “For 2025, our priority is doubling the revenue and expanding our reach in North America. Global expansion may follow, but not immediately,” Dhawan stated.

“AI-enabled services are not a fad but a reality that will transform trillion-dollar industries,” he asserted. As more firms adopt AI solutions, Attentive.ai’s early mover advantage in the landscaping and construction sectors gives it a significant edge.

Dhawan mentioned that companies such as Bluebeam, StackCT, and PlanSwift might be considered competitors, which cumulatively add to $6 million in revenues. However, they are into manual drawing tools in the construction space. For now, none offered similar services using AI, but Dhawan is sure they’ll arrive soon. 

Microsoft’s $3 billion commitment to expand the Azure infrastructure in India may have grabbed the limelight at the Microsoft AI Tour in Bengaluru; however, the company’s push to revolutionise sectors such as healthcare and agriculture also led to some key announcements by CEO Satya Nadella. 

An AI agritech startup that has partnered with Microsoft is working to eliminate guesswork in farming and empower farmers with science-backed insights to make effective decisions. 

Meet AgriPilot.ai 

Maharashtra-based AI startup AgriPilot.ai has collaborated with Microsoft Research for over five years, integrating AI, satellite imagery, and other tools to transform the farming sector. 

Identifying critical factors such as soil nutrient levels, water availability, and suitable weather conditions are some of the areas AgriPilot.ai specialises in to optimise crop yield and resource usage. 

“We have started seeing the results. Satya took that in his showcase because these are proven models now,” Prashant Mishra, founder of Click2cloud Inc., which hosts the platform AgriPilot.ai, told AIM in an exclusive interaction. Mishra confirmed that experimentation has been done for more than 2,50,000 hectares of land around the world. 

“This [AgriPilot.ai] is precisely for the marginalised farmer because Microsoft wanted to work with those with less than two acres of land. So, about a thousand farmers, with less land and resources, are currently benefitting from it,” said Mishra, emphasising that their goal is to prevent farmer suicides and distress by making them self-sustainable. 

The startup has conducted experiments, such as cultivating sugarcane thrice the size of conventional crops. It claims that the yield has doubled. Akin to conventional farming, the startup has also enabled the farming of exotic vegetables such as strawberries and dragon fruits on local farms. 

“Normally, five-star hotels import strawberries and dragon fruits from other countries, but with AI, we are able to grow them on local farms. Thanks to these exotic vegetables, the poor farmers are able to multiply their earnings, probably by 10 times or more,” he said. 

Agripilot.ai has collaborated with the Agriculture Development Trust, Baramati, which claims that these new tools have increased crop production by 20%, as presented in the Microsoft keynote session.  

AgriPilot employs a ‘no-touch’ approach, utilising satellite and drone imagery to gather farm data remotely. This allows them to provide detailed crop management plans, from pre-planting to harvesting, customised for farmers. This has all been made possible with the help of AI. 

The Microsoft Bond

AgriPilot has built a strong partnership with Microsoft Research, leveraging its advanced tools and platforms to transform farming practices. Though Microsoft has not directly invested in AgriPilot.ai, the latter depends on it for critical technological support and open-source tools. 

Nadella also met with the team at ADT Baramati, which uses AI tools to help farmers achieve healthy and sustainable harvests. 

The startup integrates Microsoft Azure Data Manager and FarmBeats to analyse soil health, monitor water availability, and optimise fertiliser usage through precise, data-driven insights. 

Empowering Women

Besides, AgriPilot partners with Pratham, a non-profit organisation, to train farmers in using these advanced technologies. This collaboration supports farmer education and provides employment opportunities for women, enabling them to conduct AI-powered soil testing independently using on-site machines.  

To ensure accessibility, instructions are provided in local languages, such as Marathi, Kannada, Hindi, Telugu, Tamil, Farsi, and Hebrew, translated through Microsoft Copilot. The collected data is then fed into AgriPilot’s platform, where AI models analyse it to deliver actionable results and outcomes for the farmers.

AgriPilot is also conducting experiments in countries like Qatar, Dubai, Peru, USA, Malaysia and India. “Baramati [Maharashtra] in India was the first [place where we experimented], and now we are doing the same in Uttar Pradesh with Microsoft,” he said. 

Big Tech’s Agri Mission

Meanwhile, other big tech companies have also been actively involved in the agri space. Last year, Google announced the availability of its Agricultural Landscape Understanding (ALU) Research API, which integrates satellite imagery with AI to deliver farm-level insights. 

The API is designed to support India’s agricultural sector by enabling data-driven decision-making, optimising farm management, and addressing productivity challenges.

Google recently partnered with the UP government to launch a Gemini-powered open network for farmers. This DPI for agriculture utilises Google’s DPI-in-a-box solution and the Beckn protocol. 

Big tech is also powering agritech startups such as Bengaluru-based Cropin, which helps farmers make informed decisions based on historical, present and future data. 

When asked about future collaborations with other companies, Mishra was clear that their focus is currently on Microsoft alone. This year, they will work towards improving accuracy before proceeding to full-scale expansion. 

Google and Meta have revolutionised marketing by turning data into performance analytics. Can something similar be done to ease the burden on teachers by making the process more data-driven? 

This idea sparked the journey of Kushal Sinha and Piyush Kumar, the founders of Chicago-based LearnQ.ai, a smart learning platform. Their efforts culminated in VEGA (Virtual Entity for Guidance and Assistance), a specialised AI agent designed to guide and assist with any task.

It represents their effort to develop a teaching assistant that can support educators and institutions, all while maintaining the core AI and data-driven approach.

In an exclusive interview with AIM, Sinha, an IIT Guwahati alumni, said, “At this stage, AI is deeply integrated into the platform, consuming vast amounts of data to provide recommendations and insights. It thereby enhances the teaching experience and empowers educators with actionable intelligence.”

The duo wanted to take a focused approach. So, instead of going too broad, they decided to test the platform on a specific use case. “We chose the SAT exam, which is widely taken in the US for undergraduate admissions. That was our initial use case. Recently, however, we’ve started allowing early customers and clients to create their own courses,” Sinha added. 

Now, they are enabling people to build courses tailored to their needs. In about a month, they’ll complete the rollout and open the platform to everyone, allowing users to build courses from scratch and deploy them.

“When we say “deploy,” the idea is to capture student interaction data—what they’re learning, where they’re struggling—and expose this data to both teachers and our AI assistant, VEGA. 

“VEGA acts as an AI tutor or avatar, leveraging the student’s knowledge graph built through interactions, assessments, quizzes, and even chat history. It also integrates the professor’s expertise,” Sinha explained. 

AI Assitant is Not AI Tutor

Kumar, the other co-founder and a former home tutor, told AIM that unlike traditional AI tutors, which often rely solely on prompts or pre-trained large models, their approach integrates real-time student data.

“For example, think of a smartwatch. If you’re self-motivated, you’ll act on its suggestions; if not, someone—like a family member—would nudge you. Learning works similarly. Some students are self-motivated, but most need additional support.”

Sinha mentioned that they use data pipelines to build a detailed knowledge graph for each student. This enables their AI system to cater responses to the student’s specific understanding level. It’s not just a wrapper around a large language model, it’s an agentic AI system that integrates data from multiple streams to deliver personalised, actionable insights.

People are now creating courses for AP, JEE, and even unique topics like teaching the Mahabharata and Ramayana. “This highlights the flexibility of our platform. Institutions like The Doon School in India are already using it to teach K-12 students,” said Sinha. Another organisation they are actively in discussions with is Allen in India. 

Why Edtechs Fail? 

“The first generation of ed-tech platforms was largely driven by the rise of the internet. Platforms like Coursera and Udemy focused on accessibility. They allowed teachers to record lectures and broadcast them widely, building platforms that connected learners to content,” Sinha mentioned. 

According to him, these platforms didn’t address efficiency. Even with all the advancements in ed-tech, students relied on tutoring services, marketplace-based learning models, and physical tuition centres. 

These platforms were more of an add-on rather than a replacement or an efficiency enhancer in education. That’s where LearnQ.ai’s approach differs, as they aim to solve for efficiency in education.

What’s Next?

Recently, Sinha and Kumar met with American entrepreneurs working with public schools in several states, who narrated a compelling success story. Four years ago, their schools ranked last in the district. 

However, by implementing a data-driven approach and using simple Excel sheets, their performance saw a sharp improvement, pulling them to the top spots in the district. Most students now score within a tight 70 to 90 range. 

That explains how any small technological adoption in the education system can bring vast improvements in a student’s performance. With AI, things are looking even better. 

In the current landscape, many AI tutors lack a robust data layer. A truly effective one must be able to understand where each student is in their learning journey and provide tailored guidance. This becomes specifically important since teachers often struggle to give personal attention to each student due to time constraints.

For instance, if you’re watching a lecture on Coursera and don’t understand a concept, there’s no way to directly ask the instructor for clarification. This is where AI avatars for teachers and virtual assistants that provide 24/7 access step in.

Another much-needed AI intervention would be a feature to transcend language barriers. For example, a teacher who speaks only Hindi and English can have an avatar that interacts in Korean or Mandarin. This would dramatically expand accessibility, allowing the same educator to engage with thousands of students simultaneously, addressing doubts and providing real-time support.

Another upcoming feature empowers educators and creators to design their own AI tools, which will align with their unique processes and expertise. For instance, a blogger could train the AI to replicate their specific writing style or workflow using a knowledge base they provide. 

These customisations open up endless possibilities for personalisation and efficiency across educational and creative fields.

The year 2024 witnessed the global adoption of an agentic AI force. From big tech companies to emerging startups, AI agents took centre stage. Keeping the momentum going, 2025 is set to be even bigger in the agentic space. A noticeable trend shows that startups building AI agents are largely being founded by Indian developers. 

The AI agents market has been witnessing promising growth. From $5.1 billion in 2024, the market is expected to hit $47.1 billion by 2030. In particular, Indian entrepreneurs have been significantly driving the growth. 

Source: X

India and AI Agents

“If you look at how LLMs have matured, it’s clear that their potential plateaus once they run out of fresh training data. AI agents pick up the slack by plugging into external sources [such as] databases, APIs, real-time feeds, and letting the model stay relevant as events unfold,” Ramprakash Ramamoorthy, director of AI research at Zoho and ManageEngine, told AIM. 

“This is why you’re seeing founders worldwide, including those from India, pushing AI agents to the forefront; it’s simply the next logical milestone for the technology after robust LLMs.”

Ramamoorthy believes that the effort to integrate AI with new data streams represents an evolution rather than a revolution and is not defined by nationality or geography but by the determination to prevent AI from becoming stagnant. “AI agents give us that infusion of timeliness and context that static LLMs alone can’t deliver.” 

Bengaluru-based AI startup Kogo AI, founded by Praveer Kochhar and Raj K Gopalakrishnan, is building AI agents and solutions to simplify workflows and improve productivity for businesses. Recently, they also launched an AI agent store. 

“We are currently building an agent that can look at a database and actually think like a data scientist or a business analyst and generate extremely intelligent questions,” Kochhar said in a recent podcast with AIM. 

The founders also emphasised the advancements in foundational models, which have made it easier and cheaper to build AI agents.

Kochhar explained how building customer support and voice-based AI agents was initially expensive, costing around ₹50 per interaction, which was unsustainable for practical use. However, advancements significantly reduced this cost to approximately ₹2.5 per interaction, making such deployments more feasible.

A number of Indian AI agents have been making the mark. A few Indian founders under the Y Combinator cohort are also building AI agents. For example, Floworks, founded by Sudipta Biswas and Sarthak Shrivastava, selected under the Winter 2023 cohort of Y Combinator, is building AI agents that will address sales functions. 

Prominent AI startups such as Sarvam AI, backed by PeakXV Partners, Khosla Ventures and CoRover, backed by Venture Catalysts, and educational institutions also have AI agents.  

AI Agents are Everywhere

With AI agents quietly becoming the norm, choosing the right sector where these agents will become the most beneficial becomes critical. “Departments such as sales, marketing, and finance usually have well-established software systems like CRM (customer relationship management), ERP (enterprise resource planning), analytics dashboards, etc., so they can plug AI agents directly into these data pipelines,” Ramamoorthy said. 

Industry leaders who have moved on to start their ventures have also gotten into AI agents. CP Gurnani, co-founder of AlonOS and former CEO of Tech Mahindra, recently spoke about how AI agents can make people more productive and efficient. “Agentic AI is the software version of a personal robot. Each one of us will have an AI agent that knows us really well,” he wrote on social media. 

AlonOS, co-founded by Gurnani and Rahul Bhatia, provides organisations with AI-as-a-service and data engineering solutions. The Singapore-headquartered AI startup recently partnered with Indosat, Indonesia’s telecom company, to accelerate AI sovereignty in their country. Though not many details have been revealed, AI agents will probably be implemented, considering that they are first catering to the travel and hospitality sector. 

Gaurav Aggarwal, who has industry experience working with NVIDIA and autonomous mobility, is now the founder and CEO of RagaAI – an AI testing platform. They recently released test frameworks for testing AI agents too. “We’re seeing AI agents evolve into so much more than just tools. They’re becoming collaborators, problem-solvers, and decision-makers,” Aggarwal said recently.  

Notably, leading database company Redis is supporting AI startups such as Kore AI in powering their virtual AI agents. “The Indian tech ecosystem is going to play a very critical role in agentic AI,” Manvinder Singh, VP of AI product management at Redis, told AIM. 

Ashwin Raguraman’s Bharat Innovation Fund began to invest in AI in 2018, starting with traditional technologies like computer vision, voice recognition, and recommendation systems. It was only in 2021 that GenAI emerged as a game-changer driven by large language models. 

This shift helped freshly minted Indian startups like Sarvam AI and Krutrim develop localised solutions. And now, it is layering into middleware (security and observability) and applications leveraging traditional and generative AI. 

So, as AI startups stagger into 2025, let’s find out just how difficult it is to build an AI startup in India. If you have a great startup idea, a business plan, and a suitable location, you just need to tap into the funds and get started. Unfortunately, it’s not as simple as it sounds. 

Stanford’s 2024 AI Index Report ranked the nations that have witnessed the most growth in AI startup activity over the last decade. According to the report, the US and China ranked at the top, while India held the seventh position. 

India’s AI Funding Game

AIM earlier reported that it is now a prime time to build an AI startup in India due to funding opportunities and acquisition potential. According to AIM Research, 43 Indian AI startups received $864 million in funding as of August 2024. Among these, Ema, an enterprise AI startup, raised $36 million in Series A funding. 

Established players such as Uniphore and Gupshup are leading the pack with late-stage funding rounds, as per Tracxn data. Uniphore raised $400 million in Series E funding (January 2022) and $140 million in Series D (November 2020). 

Similarly, Gupshup secured $240 million in Series F funding (July 2021) and multiple $100 million rounds, showcasing sustained investor confidence in AI-driven solutions.

Emerging players like Sarvam AI and Krutrim are also making waves in the industry. Sarvam AI raised $41 million in Series A funding (December 2023), signalling strong early-stage support. Krutrim, a generative AI-focused startup, has attracted $50 million in Series B funding (January 2024) and $24 million in Series A (July 2023), demonstrating consistent growth and innovation in cutting-edge AI applications.

Institutional investors such as Tiger Global Management, Lightspeed Venture Partners, and Alpha Wave Global are leading the funding of Indian AI startups. These global and angel investors are enthusiastically backing AI innovations, highlighting the country’s growing prominence in the global AI landscape.

In terms of valuations, companies receiving higher funding amounts, such as Uniphore and Gupshup, have valuations ranging from $1 billion to $2.5 billion, contributing to India’s expanding unicorn ecosystem. This reflects the increasing confidence in the potential of Indian AI companies to create large-scale global impact.

The diversity of AI applications is another hallmark of the Indian startup ecosystem. While established companies focus on conversational AI solutions, emerging players are diving into generative AI and specialised areas like text-based chatbots. For example, Senseforth is carving a niche in enterprise chatbots, while Krutrim and Sarvam AI are pioneering generative AI platforms.

Funding for AI startups in India totalled $8.2 million in the April-June 2024 quarter. In contrast, AI startups in the US received $27 billion in the same period, representing nearly half of all startup funding in the country. 

The upside? Building products in India is far less costly than in the West.

Abhijeet Kumar, CEO of Tablesprint, underscores India’s unique cost advantage. “Building a solution like Salesforce would cost millions in the US, but in India, it’s a fraction. India is no longer just a service provider; we’re creating products that compete globally. For AI startups, now is the time to build in India, with talent, resources, and cost advantages all in place.”

Are Bengaluru’s AI Startups Tempted by the Bay Area? 

For some founders, India provides what’s needed to build impactful, cost-effective technology. Amritanshu Jain, co-founder and CEO of SimpliSmart, who returned from Silicon Valley, is even more convinced. “In India, we have a deep pool of tech talent. Many think Indian engineers leave for the US due to a lack of opportunities here, but that’s changing.”

Yet, Vedant Maheshwari, CEO of Vidyo.ai, believes India’s core challenge in AI lies elsewhere. “Foundational AI requires significant capital and patience, which is harder to secure in India. While funding here is substantial, it’s mostly application-focused rather than foundational,” he explained. 

“In the US, there’s more support for deep-level work, but in India, targeting specific AI applications allows us to leverage existing models without huge initial investments.”

Vishnu Ramesh of Subtl.ai, who has ties to both the Bay Area and India, sees it as a matter of investor confidence. “The Bay Area draws investors because of its track record. Once India has its ‘Google moment’, confidence here will rise.”

Also, most IITians prefer to move to the US for better opportunities. According to the US-based National Bureau of Economic Research, one-third of those graduating from the country’s engineering schools, particularly the IITs, live abroad.

Brendan Rogers, co-founder of 2am VC, shared on LinkedIn that most of these IITians are unicorn founders. 

The Hybrid Model

India excels in application development; however, GenAI demands fresh talent and innovation. While government initiatives like the National AI Mission foster upskilling, startups continue to struggle. 

As a result, many look abroad, especially to the US, which offers faster adoption cycles, larger contract sizes, and better ROI on AI products.

As there is an evident pattern of startups moving to the US, the country continues to be a key market for Indian AI startups. It also provides higher annual contract values (ACVs), making it attractive for startups seeking rapid growth. 

Many founders adopt a hybrid model – operations and talent in India with customer bases in the US – leveraging India’s cost advantage and the SaaS model to scale globally.

Raguraman said Indian AI startups are mostly focused on enterprises rather than consumers. “The enterprise market here is an excellent testbed due to discerning customers who demand rigorous product evaluations. However, the slower adoption rates and smaller ACVs compared to the US remain a challenge. This disparity often compels startups to focus on international markets for growth while maintaining a foothold in India,” he said.

What’s Next?

What venture capitalists look for while evaluating AI startups is, how much effort they have invested in their technology, what proprietary data they control, and the unique value they’re adding over existing models. 

This effort should be substantial, so they are confident it’s not just a surface-level improvement but something with real depth.

Indeed, India offers a more capital-efficient startup environment, whether for AI, or otherwise. However, the question is whether this will remain the case as these businesses scale globally. Once startups expand and start competing internationally, they will need to invest in talent from around the world, which could increase their expenses.

In terms of capital, India requires less funding for startups compared to the US, but there is also significantly less capital available here. While the Indian VC and startup ecosystem has grown substantially, it’s still relatively young and about 60 to 70 years behind the US ecosystem. 

Critical care in India faces a major crunch, with estimates suggesting only 2.3 ICU beds per 100,000 population. Furthermore, intensivists or critical care doctors are also in short supply, with only 5,000-6,000 trained professionals in our country. This shortage becomes a bigger threat in smaller towns and non-metro regions, leading to unfortunate, preventable deaths. 

A Bengaluru-based healthcare startup, Cloudphysician, aims to address this disparity with AI. 

ICU Care 2.0

Cloudphysician was founded in 2017 by Dileep Raman and Dhruv Joshi, two US board-certified intensivists who have witnessed the healthcare system in the West. They built the platform with a mission to use AI and telemedicine to bridge the skill and resource gap in India’s critical care infrastructure. 

By connecting ICUs through high-quality video and bedside data analytics, Cloudphysician looks to improve patient outcomes in both neonatal and adult critical care. 

“We have approximately 3.5 lakh ICU beds in the country. However, for a country our size, we need between 8 to 10 lakh ICU beds,” said Raman in an exclusive interaction with AIM. 

“It’s not that you put a bed and a ventilator and add some devices, and it becomes an ICU bed. Besides the must-have hardware and the infrastructure, you also need skilled people to run it and a set of processes that make the high-quality ICU function. That’s what makes an ICU bed,” explained Raman.  

Healthcare Operational AI

Cloudphysician uses a combination of multimodal AI models, incorporating inputs from video feeds, lab results, medical records, ambient audio, and established medical guidelines. This integrated approach allows the AI to detect critical issues, such as potential infections or tube disconnections, and provide actionable insights to doctors in real time. 

“So it is not about predicting who is going to get worse or better. It’s more about analysing what exactly is going on because that is what enhances the efficiency of the doctor with us,” explained Raman.

They use a combination of computer vision models for visual analysis and LLMs for reasoning and recommendations. Raman said they also leverage platforms like Google Cloud and OpenAI alongside their in-house models. 

The startup currently covers over 1,500 ICU beds across 200 hospitals in more than 100 cities throughout India and has even demonstrated a significant impact on reducing mortality by up to 47% in certain ICUs. 

A few months ago, the startup raised $10.5 million in a funding round led by PeakXV Partners, Elevar Equity and Panthera Peak. 

Humans in the Loop

The platform extensively employs AI, albeit as an augmenting tool. “The AI is not making any patient care decisions. It’s still the doctor and the nurse, but they’re doing it in a far more efficient manner now,” said Raman. If an ICU doctor can see 8-10 patients, Cloudphysician will be able to increase that by 6-8 times. 

Hands-on clinical training is a requirement for the startup, so half of its workforce consists of clinicians, doctors, and nurses who undergo intense training. Currently, the startup has a 280-member team, and HCG, Motherhood, and Cytecare cancer hospitals are some of its customers.

AI in Healthcare

A number of AI-based health tech startups have emerged recently, with the goal of addressing the critical staff shortage in patient care. The AI in healthcare market is expected to grow to $173.55 billion by 2029. 

Dozee, a Bengaluru-based startup, offers an AI-based contactless remote patient monitoring system. It tracks key metrics such as alert sensitivity, specificity, response time, and healthcare activity. The system is said to have the potential to save 21 lakh lives annually and reduce healthcare costs by INR 6,400 crore.

Cloudphysician has an ambitious vision to become the global engine for delivering healthcare, “Just like what happened in IT services decades ago,” remarked Raman. 

In India, cancer treatment often follows a standardised approach, bucketing the treatments based on the stage and type of cancer. However, the West adopts a rather nuanced strategy with genome research to understand the exact type of cancer within a certain stage and type and administer targeted therapy, thereby hitting a higher recovery rate. 

Hitesh Goswami and Kshitij Rishi founded the Bengaluru-based oncology precision startup 4baseCare in 2018 to introduce a similar approach to treating oncology patients in India. Recently, Infosys Innovation Fund invested close to $1 million in the leading precision oncology firm. They were also the past winners of the Karnataka government’s ‘Elevate’ initiative that supports innovative early-stage startups. 

A few months ago, deep tech-focused venture fund Yali Capital led a Series A funding round and raised $6 million.

Genome Technology for Targeted Care

In an exclusive interaction with AIM, co-founder and CEO of 4baseCare, Goswami, explained the methodology behind cancer treatment. Previously, cancer treatment was uniform. All patients received the same therapy based on the cancer type and stage, and individual differences were not considered in treatment plans.

Goswami said that today, lung cancer stage-2 and stage-3 patients are subgrouped into 12 to 15 categories; each requiring distinct treatments as therapies effective for one group may not work for another and could even cause adverse reactions.

According to him, sequencing one human genome took around 15 years and $3.2 billion. Now that the Human Genome Project has concluded, Goswami believes the procedure can be done at a cost of $100 for a genome in a couple of years.

Notably, genome testing for oncology has grown rapidly in India. What stood at 5,000-6,000 tests in a year in 2019 has now touched 2 lakh. Goswami estimates that the market will easily hit 3 lakh tests.  

India has its Own Struggles

While the results help in targeted care for cancer patients, precision oncology in India faces key challenges, including limited awareness among oncologists in tier-2 and tier-3 cities, who often rely on traditional chemotherapy. Many patients are also unaware of advanced targeted therapies. 

Further, when effective drugs are identified, they are often unavailable in India. This becomes a cause for frustration among both patients and doctors. Affordability also remains a key challenge as treatments like immunotherapy cost ₹2.5-₹3 lakh per cycle and require multiple rounds. 

“Although many pharma companies are running a lot of patient support programs where they give free access to a certain extent to some drugs. But overall…it is very expensive,” Goswami pointed out. 

With increased awareness and reduced costs, the method is expected to become more accessible and be adopted in the coming years. 

Analysis of Gene Fusions across functional categories and cancer types using 4baseCare’s study. Source: 4baseCare

AI at Play 

4baseCare has been adopting AI for a number of its use cases, including data interpretation, where it uses AI models to get the right insights from the huge amount of data it generates. 

“So, right now, we are telling patients and doctors that you can look at your applications and see what you can work on. The next level that we are working on is directly giving the recommendation, best recommendation, best three recommendations in terms of application,” Goswami explained. 

The startup has also received a grant from the Indian government to develop 3D cell models of tumour tissues in a lab setting. These models are used to test the effects of various drugs on the tumour, and the data collected from these experiments is being used to train an AI model to predict which treatments are most effective for specific types of patients.

In addition to this, the startup is working on a novel concept referred to as the CG Twin, which stands for clinical genomic digital twin of cancer patients. The method involves integrating a patient’s clinical and genomic data into a comprehensive profile. This system will allow doctors to compare a new patient’s profile with a database of similar patients, known as twins, and by analysing these matches, doctors can access insights about similar patients’ treatments, outcomes, and risk factors, thereby enabling a more personalised and informed treatment plan based on real-world data and previous cases. 

“We have done close to 15,000 plus tests now, and we have actually developed a machine learning algorithm, which has been trained in 15,000 patients to build this twin model,” said Goswami, who looks to build a future which will be more evidence-based and outcome-based decisions rather than an empirical way of providing treatment. 

CG Twin is already being tested by doctors and is slated to be released in the upcoming months. 

Notably, the concept of digital twins in healthcare is gaining traction. Several healthcare platforms have used NVIDIA’s Omniverse to build a simulated environment of patients to help understand and administer more efficient treatment. 

The Mission Continues Amidst Struggles

Goswami explained that the unique name of the startup is inspired by the four bases of DNA – adenine (A), cytosine (C), guanine (G), and thymine (T) – and the four pillars of cancer care – allied care, technology, global genomic research, and clinical care.

“What we realised is that these four bases or these four pillars are working in silos. So, we wanted as a company to bring all these four bases together just like the four bases of DNA to provide cancer care,” he said.

Notably, 4baseCare has partnered with well-known hospital chains such as Apollo Hospitals, Fortis Healthcare and Tata Memorial Centre.

The journey has not been an easy one, with the startup having faced a fair amount of funding struggles. “When you’re talking about genomics, genomics-driven data and all, there was a lot of apprehension because genomics is something new. And, it is something that’s not easy for everyone to understand.” 

Goswami even recounted how they had to withstand “a hundred noes” before one “yes”, something that is a common reality for many founders in deep tech. However, eventually, things have a way of falling into place. “You need that one yes from the right people to believe in your vision, and I think that’s what we got with Yali and with Infosys.”

Above all, the biggest motivation for Goswami and his team, which has close to 200 members, comes from the profound impact their work has on cancer patients.

Goswami believes that seeing patients, once suffering from salivary gland cancer, cancer-free because of their recommendations is what keeps them motivated.

“The whole team, right from the logistics team who picks up a sample, I tell them, ‘Guys, you have no idea which sample you might pick up, and that will change the whole family’s life. So right from every level, someone is somehow impacting a family which they don’t even know. So that has kept us going, and…it’s a very exciting journey,” Goswami concluded. 

Meta AI chief Yann LeCun, in a recent interview with Zerodha co-founder Nikhil Kamath, advised budding AI entrepreneurs in India to pursue an academic degree, such as a master’s or PhD, particularly in technical and complex fields like artificial intelligence, before building a startup.

“Doing a PhD or graduate studies trains you to invent new things and ensures that your methodology prevents you from fooling yourself into thinking you’re being an innovator when you’re not,” he said.

LeCun added that while a PhD is not a strict requirement for success, it offers significant advantages for entrepreneurs. “It gives you a different perspective,” he said. “In a complex, deeply technical area like AI, it’s useful to learn about what exists out there, what’s possible, and what’s not.”

His comment resonated with many founders. “If you are building a deep-tech startup, which is more than just a GPT wrapper, you will need technical people in the core AI team,” said Pijush Bhuyan, computer vision engineer at Awiros. He added that companies need people who have got their hands dirty with PyTorch, and who have spent hours implementing state-of-the-art research papers, not prompt engineers or people hailing from an SDE background without prior experience.

Amit Sheth, the chair and founding director of the Artificial Intelligence Institute at the University of South Carolina (AIISC), agreed with LeCun, saying that this was definitely true for deep-tech startups. “Several of my PhDs have started successful startups (and I have done four). Of course, there are plenty of cases where founders did not get advanced degrees but have succeeded, so this is unnecessary, but in general, those with deeper experience in technical innovation have an edge,” he said.

Startups founded by Sheth’s students include AppZen, thunk.ai, Objective, Inc., and Clinical AI Assistance.

Rajan Anandan, managing director at Peak XV Partners (formerly Sequoia Capital India), believes the Indian startup ecosystem would benefit from more founders with deep expertise in artificial intelligence and software development. According to a research from Private Circle, only 8% of newly launched AI startups have founders with a PhD degree.

A 2017 study by the National Bureau of Economic Research showed that startups with at least one PhD founder are more likely to succeed and have higher valuations at IPO. The research found that having a PhD founder increases a startup’s chances of a successful exit by over 50%.

Vishnu Vardhan, the founder of Vizzhy and SML, aptly described the sad state of Indian AI startups in an exclusive interview with AIM, stating, “Indian startups often focus on business applications instead of foundational innovation, with investors prioritising quick returns over long-term deep-tech investments.”

He said that the so-called deep tech investors clearly have no theses whatsoever. “I met a few VCs who said they were deep tech investors. I asked them their ticket size; they don’t even understand the scale of investment required for true deep tech.”

Exceptions Galore 

Elon Musk, the man behind companies like Tesla and SpaceX, never pursued a PhD or master’s degree. Despite this, he successfully built and led some of the most innovative companies in the world.

Interestingly, earlier this year, LeCun engaged in a banter with Musk, who questioned the former’s contribution to AI by asking how much research he had conducted “in the last five years.” Pat came LeCun’s reply: “Over 80 technical papers published since January 2022.”

“SpaceX would not exist without the thousands of scientific papers on rocket engine design, propellant chemistry, rocket control, material science, orbital mechanics, heat dissipation, trajectory planning, and the hundreds of scientists who got where they are by studying these papers,” claimed LeCun.

Deep Tech Startups in India 

Similarly, Vishnu Ramesh, founder of Subtl.ai, couldn’t agree more. “We have come so far with Subtl.ai thanks to people like Manish and Pranav coming in as my partners. Super excited to disrupt the custom SLM RAG space!” he said.

Manish Shrivastava, co-founder and chief scientist at Subtl.ai, has a PhD in computer science from IIT Bombay. Similarly, CTO Pranav Goyal has a dual degree, a BTech in computer science and an MS in computational linguistics from the International Institute of Information Technology, Hyderabad (IIITH).

Notably, Sarvam AI, one of the most popular AI startups in India, was founded by Vivek Raghavan, who has a PhD in electrical and computer engineering from Carnegie Mellon University. His co-founder, Pratyush Kumar, also holds a PhD from ETH Zurich.

Similarly, Pranav Mistry, founder of TWO AI, completed his BE in computer science from Gujarat University, followed by an MDes in Design from IIT Bombay. Ritwika Chowdhury, the founder of Unscript, holds an MTech in electronics and electrical communication engineering from IIT Kharagpur.

Rise of AI Startups by Researchers

In the West, many AI researchers have recently founded their own AI startups. François Chollet, the creator of Keras, recently announced his departure from Google. Not long after, Toby Shevlane, a scientist at Google DeepMind, also revealed that he was leaving the company to pursue his own venture.

Nearly all researchers who co-authored Google’s Transformers paper ‘Attention Is All You Need’, which shaped the foundation of modern AI, later left the organisation to start their own companies to tackle specialised niches within AI.  

“…At that time GPT-2 had just come out and the trajectory of the technology was pretty clear…So I called up my co-founders and I said ‘Maybe we should figure out how to build these things,’” said Cohere CEO Aidan Gomez in a recent podcast, elaborating on the need to capitalise on the wave of future internet models. 

Fei-Fei Li, co-director of Stanford’s Human-Centered AI Institute, co-founded World Labs in 2024. The institute, valued at over $1 billion, aims to develop AI systems with advanced spatial intelligence for 3D interaction. 

Similarly, Ilya Sutskever, co-creator of ChatGPT, shifted his focus to AI safety by founding Safe Superintelligence, reportedly valued at $5 billion. Meanwhile, in March, computer scientist Kai-Fu Lee founded 01.AI, valued at $1 billion, to develop open-source LLMs specific to China.

No-code platforms are not dead yet. They are very much alive with the help of AI. Bengaluru-based Tablesprint aims to stand out as an AI-powered low-code platform, designed to bring enterprise-level applications to companies of all sizes. 

Following the recent $1 million seed funding round led by Ola’s co-founder Ankit Bhati and other notable investors, Tablesprint co-founder and CEO Abhijeet Kumar discussed with AIM the company’s vision, its distinct position in the market, and its potential to transform business operations globally, competing with the likes of Salesforce.

Kumar’s background includes founding and scaling a successful startup that eventually became BigBasket’s backbone. His experience in handling operations-heavy systems provided the foundation for Tablesprint’s solutions. “BigBasket is an enormous operation, especially from an operational and logistics perspective,” Kumar noted. 

“It’s a consumer tech business that relies heavily on robust technology. That experience equipped us with the knowledge to tackle complex tech challenges, which now fuels Tablesprint’s mission.”

Another Low-code/No-code?

While many platforms provide low-code options, Tablesprint aims to address a crucial gap in enterprise technology: accessibility and control over data. Traditional low-code platforms often offer limited control over sensitive information due to rigid structures and limited permissions. 

Tablesprint’s platform addresses this by incorporating robust data controls, permissions, and a secure backend infrastructure that integrates seamlessly with AI, ensuring that businesses can safely and effectively leverage their data. Rather than attempting to build AI models from scratch, Tablesprint focuses on integrating and optimising existing open-source and commercial models to provide a comprehensive, end-to-end solution that enterprises can trust.

The moat of Tablesprint? Unique blend of affordability and flexibility, positioning it as an alternative to established players like Salesforce and Zoho. “We’re offering the flexibility of Salesforce but at a Zoho price,” Kumar explained. “Our focus is on creating a customisable, scalable system that can handle complex enterprise needs without the high costs usually associated with such solutions.”

Tablesprint’s competitive advantage, Kumar emphasised, lies in its capacity to offer a Salesforce-like experience at a fraction of the cost, with a user interface (UI) that simplifies complex tasks. “With us, you’re not just getting a software tool; you’re gaining a system that adapts to your needs, allows for high-level customisations, and integrates AI capabilities directly into the workflow.”

Chirag Jadhav, co-founder and CTO, earlier highlighted the platform’s multi-tenant system designed for both developers and business users. “Creating a system that resonates with both developers and business stakeholders is a significant challenge. We are excited with our progress so far and look forward to tackling more challenges ahead on this journey.”

Founded earlier this year, Tablesprint’s AI-first SaaS platform allows companies to rapidly build customisable applications for various business functions, including HR, sales, operations, and vendor management, using open-source and closed-source models available in the market. 

Its no-code solution features modular building blocks such as AI write/image tools, workflows, and charts, enabling businesses to start with simple tasks like surveys and scale to full-fledged workflows.

Staying Alive and Sprinting Ahead

Kumar is aware of the competition from established giants like Salesforce and Zoho, but he believes Tablesprint offers unique advantages that make it attractive, particularly for businesses conscious of both costs and customisation capabilities. 

“Try making a form in Zoho, and you’ll see it’s not designed with seamless user experience in mind,” he said. “Our UI is optimised to make processes intuitive and efficient, making it appealing to users who need functionality without the overhead.”

In addition, Kumar highlighted the often limited adoption rate of Salesforce due to its complexity and high price point, particularly in markets like India, where cost sensitivity is paramount. According to him, only around 34% of Salesforce licences end up being used because companies struggle with its complexity. “We’re targeting that space by offering a similarly robust platform that’s also simpler to use and much more affordable.”

The company’s go-to-market strategy involves partnering with implementation providers, giving entrepreneurs and developers the flexibility to use Tablesprint to build apps, which they can then offer to clients as solutions. “Tablesprint is not just an India-specific product; it’s a global solution designed to help businesses worldwide streamline operations, build apps, and generate revenue by offering our platform as a customisable solution,” Kumar explained.

Tablesprint’s roadmap is ambitious. The company plans to expand its offerings by building a complete AI-based ERP system and form builder, with the goal of becoming a one-stop shop for enterprise applications across industries. 

Despite the increasing saturation in the low-code, no-code market, Kumar is confident that Tablesprint’s model and global reach will help the company thrive in the long term.

“We’re still a startup, so bandwidth is sometimes an issue,” he admitted. “We have so much interest that we have to pick and choose our clients carefully because we want to ensure each implementation is successful and adds value.”

Floworks, the cloud-based enterprise automation startup, recently released a novel ThorV2 architecture allowing LLMs to perform functions with better accuracy and reliability. The YC-backed company collaborated with IIT Bombay and Kharagpur to build this architecture. 

In an interview with AIM earlier this year, Floworks claimed that its AI agent, Alisha, is 100% reliable for tasks involving API calls. Sudipta Biswas, the co-founder of Floworks, said, “Our model, which we are internally calling ThorV2, is the most accurate and the most reliable model out there in the world when it comes to using external tools right now.”

Further, he claimed that ThorV2 was 36% more accurate than OpenAI’s GPT-4o, 4x cheaper, and almost 30% faster in terms of latency.

These sensational claims were recently backed by an in-depth research paper that dives deep into ThorV2 architecture and how all of its novel features work to solve several crucial challenges in agentic workflows inside some of the market-leading LLMs today. 

Edge of Domain Modeling is ThorV2’s Hero Technique

The edge of domain modelling, used in ThorV2 architecture, involves providing minimal instruction upfront, allowing the agent to begin the task, and then providing the remaining information through error corrections post-task.

This approach differs from providing knowledge of all possible scenarios regarding the function calling. 

Edge of domain modelling reduces the need for extensive instructions, which in turn reduces the number of tokens in the prompt. It can further lead to cost saving measures. 

The authors mentioned that “Function schemas can be lengthy, leading to large prompt sizes. This increases deployment costs, time consumption, and can result in decreased accuracy on reasoning tasks.” 

Additional instructions added to the LLM in the error correction process are performed by a static agent implemented through an Agent Validator Architecture inside ThorV2. 

You Don’t Need an LLM to Evaluate Another LLM 

The Agent Validator architecture overcomes several limitations of agentic workflows, where the primary LLM agent performing a task receives feedback from other LLMs that act as critics.

The authors argue that using an additional LLM not just increases the deployment costs, but decreases the rates of accuracy. 

The introduction of a static agent written fully in code, includes a component called Domain Expert Validator (DEV) which inspects the output generated by the LLM for errors. The DEV contains all the knowledge and information required to perform the function calls inside a specific platform.

While building a validator requires a significant amount of effort, it helps reduce processing time and improve accuracy. This is because the knowledge contained inside the DEV contains information regarding the most common and repetitive errors that occur in the function calling process. 

Multiple API Functions in a Single-Step 

One of ThorV2’s other advantages is that it can generate multiple API calls in a single step. With ThorV2, a single query is sufficient for both tasks, even if the first task needs to retrieve information from the API for use in the second task.

The approach involves using a placeholder to represent unknown values, and once the first task retrieves the API response, the value is then injected into the second task.

“Generating multiple API calls at once requires sophisticated planning and reasoning capabilities, which is very challenging for ordinary LLMs. Our Agent-Validator architecture simplifies this process as well by correcting errors in the planning step”, added the researchers. 

This approach is a significant improvement over the traditional, sequential handling of API calls in current LLMs, which often require a step-by-step execution process.

And the Numbers Don’t Lie – 50% Cost Reduction With 100% Reliability

The ThorV2 architecture was compared to OpenAI’s GPT 4o, GPT 4 Turbo, and Claude 3 Opus for a set of operations on HubSpot’s CRM.

The authors developed a dataset called HubBench on which the model was evaluated. The models were tested for accuracy, reliability, speed, and cost. In a conversation with AIM, Sudipta mentioned that ThorV2 was connected to the Llama 3 70B model for comparison. 

ThorV2 came out on top in every single test, and a 100% score in the reliability test, which seeks a consistent output when the model is put out to perform the task ten times.

In the single API call function, ThorV2 scored 90% accuracy, second to Claude 3 Opus’ 78% score.

The test also revealed that it only took $1.6 for a thousand queries, which is 3 times cheaper than OpenAI’s models. Even with multiple API calls, ThorV2 performed better on every single metric. 

While reading the comparison benchmark scores, one wonders if these scores are relevant five months after the tests were conducted, with several new and capable models like Claude 3.5 Sonnet and GPT o1 having been launched.

However, it is important to understand that ThorV2 is an architecture built to enhance the performance and capabilities of an existing LLM. The integration will, in fact, work better with new and more capable models. 

“We will soon come up with Thor v3, which will definitely compare with other models that have come up recently. But again, the framework is not a model-level innovation that we’re doing,” Sudipta said.  “So even if the underlying model keeps on getting better, our framework will keep supporting even better than that.” 

It Isn’t Perfect, But Floworks Wants to Get There 

One of ThorV2’s limitations is that it relies on knowledge from the DEV based on common, and well established error patterns and it may face difficulties approaching an unseen one. Moreover, the research currently tests ThorV2’s architecture for just single, and two API call functions.

The authors acknowledge the limitations, and plan to perform a comparison with three or more function calls in future research. 

In the conversation with AIM, Sudipta revealed that ThorV3 is currently in the works, and it will challenge some of the latest market leading models today. That said, one can also expect other limitations to be resolved in the future iteration. 

A Vision to Solve More Real-World Problems

The authors envision ThorV2 to overcome the limitations of existing LLMs and solve problems that can truly create an impact. 

They mentioned that LLMs have revolutionised NLP and AI, demonstrating remarkable capabilities across a wide range of tasks. However, their economic impact has been somewhat limited, particularly in domains requiring precise interaction with external tools and APIs.

Over the last few months, we’ve also seen a meteoric rise in AI Agents and their tremendous capabilities, and frameworks like ThorV2 can only propel their powers further in sectors that require a large amount of automation and knowledge transfer between different applications. 

“LLMs seem very cool, but to front-load them with a high amount of tokens, the cost will be prohibitively, very high. For large-scale operations where lots of automation is needed to be done, that price point will not suit enterprises, and small businesses,” Sudipta said.