Bhavish Aggarwal is not a person who does things incrementally. When he built Ola, he did not try to improve on existing taxi services — he reimagined urban mobility for a country where most people had never taken a taxi. When he started Ola Electric, he bet that India would leapfrog the internal combustion engine entirely. Krutrim follows the same pattern: not a better version of existing AI, but a fundamentally different approach built for a fundamentally different context.
The Frustration That Started Everything
The origin story of Krutrim is a story of frustration. As Ola grew into one of India's largest technology companies, Aggarwal found himself increasingly dependent on AI tools that were not built for India. The language models that powered customer service were trained primarily on English data and struggled with the Hindi, Tamil, and Telugu that most Ola customers and drivers actually spoke. "I kept asking our teams to use AI to solve problems, and they kept coming back and saying the tools do not work for our context," Aggarwal says. "At some point I realised that the tools were not going to get better for India unless someone built them specifically for India."
The Full Stack Vision
What distinguishes Krutrim from other Indian AI companies is the ambition of its scope. Most AI startups focus on a specific layer of the stack — a foundation model, an application, a vertical solution. Krutrim's stated goal is to own the entire stack, from custom AI chips to consumer applications. The chip ambition is the most audacious element of this vision. Aggarwal argues that India's dependence on foreign chip suppliers is a strategic vulnerability that will become more acute as AI becomes more central to the economy.
The Language Model Architecture
Krutrim's foundation models are trained on what the company claims is the largest corpus of Indic language text ever assembled for AI training. The model architecture incorporates several innovations designed specifically for Indic languages. The tokenisation system handles the morphological complexity of Indian languages more efficiently than standard tokenisers. The code-switching capability — the ability to handle sentences that mix multiple languages, as Indian speakers naturally do — is a particular point of pride. "Real Indian communication is not monolingual," Aggarwal says. "A business meeting in Bangalore might involve English, Hindi, and Kannada in the same conversation. Our models need to handle that naturally."
The Consumer Product
The Krutrim assistant, which launched in early 2024, is designed for voice-first interaction — a deliberate choice that reflects Aggarwal's belief that voice, not text, is the natural interface for the hundreds of millions of Indians who are comfortable with smartphones but not with typing. The product has accumulated millions of users, with particularly strong adoption in Tier 2 and Tier 3 cities. "The people who benefit most from AI are not the people who already have access to everything," he says. "They are the people who have been excluded from the benefits of technology because the technology was not built for them."
The Unicorn Milestone and Beyond
The $1 billion valuation that made Krutrim India's first AI unicorn was, in Aggarwal's characteristically understated framing, "a good start." The roadmap ahead is ambitious and long. The chip design programme will take years to produce results. The enterprise AI business requires building sales and support capabilities that are still nascent. "We are building something that has never been built before," Aggarwal says. "A complete AI stack, built in India, for India, that can eventually compete globally. That is a ten-year project, not a two-year project."