Mine began with reach. I joined IIT (BHU) for Electronics Engineering in the late 1990s, right on the cusp of India’s internet boom, and what excited me then was the same thing that excites me now: that a piece of code can reach millions of people and meaningfully change lakhs of lives. Two decades of enterprise work across Microsoft, Expedia, IBM, and Mentor Graphics gave me scale and craft, but the spark from those early days had quietly dimmed. I had been tracking machine learning since 2008 and got serious about deep neural networks after reading about GPT in 2019. An advanced AI program at IIT Kanpur instilled in me the rigor to train models from first principles. By 2024, when I saw how badly modern AI systems were failing Indian languages on the Bhashini dashboard, I knew where my next twenty years had to go. Within months, I had left Microsoft to start DeepMindSecure.

The inspiration came from two places. The first was an Economist article in 2024, noting that Indians were among the heaviest users of voice as an interface with AI, and that the AI was failing them. The second was personal. I had the privilege of an English-medium education; many of my closest childhood friends did not. I watched them prepare for competitive examinations and navigate their early careers in Hindi, Bengali, Tamil, Telugu, and Marathi. I watched how often that linguistic boundary became a career boundary. That should not be the case in our own country. Voice is the most natural interface humans have, and India is fundamentally an oral civilization. If we build Voice AI that truly understands how Indians speak, accent, dialect, code-mixing, all of it, we don’t just build a product. We open a door.

Each phase taught me something different. Mentor Graphics gave me engineering depth, the discipline of building systems where correctness is non-negotiable. IBM taught me how enterprise software is sold, supported, and made dependable across geographies. Expedia put me close to consumer products at scale, where milliseconds and edge cases decide whether a user returns. Microsoft was where I saw the AI shift up close and learned what it takes to quietly build cloud-grade infrastructure that millions rely on. The common thread across all of it was this: great engineering organizations are built on clarity of mission and the courage to make architectural decisions early, before the cost of changing them becomes prohibitive. That single lesson shapes how we run DeepMindSecure today: a small team, a sharp thesis, and full-stack ownership.

I think about this through a historical lens. Our colonial heritage kept India out of the Industrial Revolution, and we are still, in the 2020s, trying to catch up on manufacturing. AI is the equivalent of an inflection point for our generation. If we do not build the Indian AI stack, models, infrastructure, and the services layered on top, we will spend the next hundred years renting capability from someone else, at someone else’s prices, under someone else’s terms. Sovereignty here is not a slogan; it is economics, security, and cultural fidelity. A foreign model trained predominantly on Western data will never understand a Haryanvi grandmother the way a model built in India for India can. We need a few frontier models out of India, but the real public benefit will come from many small, purpose-built models that ordinary Indians can actually afford to use.

Education, first and decisively. That is where the English-medium barrier first surfaces for a child, long before it appears in a job interview or a hospital queue. Asking a child to learn mathematics and science in a language they do not yet think in is asking them to do two hard things at once. India-focused Voice AI can change that classroom dynamic entirely.

BFSI is the second and has the most organic pull. Jan Dhan and Aadhaar have brought hundreds of millions of Indians into the formal economy, yet barely five percent regularly use e-banking. The friction is sign-in complexity and a genuine fear of online fraud. Indians today are largely consumers of online content rather than participants in online services. Accurate, secure Voice AI in the customer’s own language is what closes that gap.

Agriculture follows, a sector we are not directly building for, but one where small purpose-built models fit beautifully: a farmer photographing his field, getting a yield estimate, and acting on advice in his own tongue, even as climate change pressures long-term output. The list of India-specific domain problems that small models can solve is endless. That, ultimately, is where AI’s maximum impact will come from.

The first challenge is mindset. Most modern AI systems treat Indian languages as a single statistical bucket and dialects as noise. The truth is the opposite: a person in Mainpuri speaks Kannauji, not “Hindi,” and a model that does not respect that distinction will fail them. At Sarvbol, we treat dialects as first-class citizens. The second challenge is data. People assume more data is always better. For Small Language Models, variety and coverage matter far more than volume; you need every phonetic combination, every age group, every gender, every social setting represented. We ran a scientific voice data collection effort with proper population segmentation across Awadhi, Kannauji, Braj, Bundeli, Haryanvi, and Bhojpuri to ensure coverage. The third challenge is benchmarking; Indian benchmarks like Graamvaani are still maturing, and you have to push them forward even as you build against them.

The honest answer is restraint. The temptation in AI is to chase the largest model, the most parameters, the most data. We deliberately went the other way. Sarvbol is built as a family of small, purpose-built models that run on mid-sized GPUs, rather than the expensive data-center-grade hardware that makes AI unit economics painful in India. We own the full stack: hardware choices, data collection, model architecture, and inference. That ownership is what makes scale possible at Indian price points. A wrapper over a foreign API can never be cheap enough, fast enough, or compliant enough for true mass deployment in India. Purpose-built, sovereign, and economically rational, that is the only recipe I have seen work for AI products in this market.

It is foundational, and it will only become more so. Enterprises and government bodies cannot adopt AI at scale if they cannot verify where their data lives, who can see it, and how the model uses it. This is why on-premise and sovereign deployment is not a nice-to-have for India; it is the unlock. One of our six patent filings, and the one already granted, is in the area of on-premise secure inference. It allows customers in regulated industries, healthcare, BFSI, and defense-adjacent to run powerful AI inside their own infrastructure without compromising compliance posture. In our view, trust is not a marketing layer applied at the end. It is an architectural decision made at the beginning. The companies that get this right will define the next decade of Indian enterprise AI. The ones that don’t will struggle to move beyond pilots.

Three principles. First, clarity of mission above all, engineers do their best work when they understand not just what they are building but why it matters and who it serves. Sarvbol’s mission of linguistic inclusion is something every member of our team carries with them. Second, technical depth at the top, I do not believe in leaders who have drifted too far from the code. In AI especially, architectural choices must be made by people who can still reason about gradients and inference paths. Third, ownership over hierarchy. Small teams that own outcomes end-to-end will outperform large teams with handoffs every time. Our patents, our benchmark performance, our pace, none of it would have been possible without that culture of compact, deeply accountable ownership.

I see Voice becoming the default interface for the next billion Indians coming online, not the keyboard, not the touchscreen menu, not the chatbot in English. A farmer asking about crop advisory in Bhojpuri, a patient describing symptoms in Awadhi, a citizen filing a grievance in Haryanvi, a student preparing for examinations in Marathi, every one of these interactions should be effortless, and today they are not. Voice AI that genuinely understands Indian speech is not a feature. It is a civic infrastructure layer. I am an AI optimist. I believe AI will be a force of inclusion in India, not a widener of the have-and-have-not divide, but only if we build it ourselves, in our languages, on our terms. That is what Sarvbol exists to do, and that is what the next decade of our work is about.