Jan Ki Baat — Scaling voice AI for India's public infrastructure
Four days before our first big demo, every volunteer on the project had quit and the prototype was broken. I spent a Saturday at Tanuj's house, threw the whole thing out, and rebuilt it from scratch in a single day.
Jan Ki Baat is a voice agent that calls people and listens, in their own language, so government can actually hear them. I started it at People+AI, ran it for fifteen months, and handed it on when it outgrew me. I had never touched voice before.
where it started
At People+AI we were not handed a project. The first job was to go find a problem worth solving. We found this one: most people in India have no real way to reach their government, and the government has no real way to hear them back. I knew the feeling first hand.
The question was whether you could stitch a few pieces into a voice agent that just calls people and listens, in their own language, at scale. This was early. There was barely any voice telephony to build on. I believed it would work and figured I would keep iterating until it did.
the four-day rescue
The demo was four days out. Every volunteer on it had left, and the prototype one of them had built did not work and was full of holes. So on a Saturday I went to Tanuj's house, sat there the whole day, threw out the broken version, and built it again from scratch. Zero to one, in a day. I was scared the whole time.
I walked out with a working prototype. There was no one left to save the project, and I did. It is the thing I am proudest of.
getting the go
The first time I showed it to Shankar Maruwada, who runs EkStep, I was in the room with Tanuj and Arun. He looked at it and told me this would be the biggest project of my life, and to take a photo because it would be useful one day. He was right about the project. We never took the photo.
No one told me it was a bad idea. I was not nervous. I was excited. I already knew.
the Nilekani demo
What we built first was the voice agent plus a tool that turned every call into a readable insight, so you could actually see what people said. In February we showed it to Nandan Nilekani. The Sarvam founders were building a version of the same thing, so we presented together.
Pratyush showed his first, a voice demo, and it glitched on the day. That is the day I learned the phrase "demo gods." Mine held up, and I showed the metrics. We left that room having decided to do a big launch, around the ninth of March.
fifteen months
My job was to keep things moving and get things done. I would hire a volunteer, hand them work, check if it worked, and when it did not, build it myself and ship it. To validate the idea we ran a cohort, invited around thirty-five organisations, and picked a set of them. To run all of it I put out an open call for product managers, program managers, and engineers, interviewed people, and built a small, sharp team that carried the whole thing.
I was assigning projects, tracking them, building, and throwing in new ideas at the same time: anti-spam, a protocol to cut cost, an effort to train our own Indic speech models, and working the demand and supply sides by talking to early voice startups.
The best thing that came out of that cohort was not a number. Three or four voice startups grew out of that group. People who volunteered with me went and started their own companies. One of them started Dialflo, where I am now.
pilots and adoption
We did not build tech in-house, on principle, so half my job was finding vendors who could build, handing them clients, and testing what came back. The first real pilot we sponsored ourselves: two hundred calls for a nonprofit in 2024. Good results. A small number, but it was the proof we needed.
From there it moved. A friend working with Arghyam liked it, and we took it to a state, where it became a state-level adoption. We sent recordings to the Prime Minister's Office. I reached the IT Minister of Telangana. With ONEST I ran the DePwD pitch. The official there wanted to bring AI in during his last stretch in the role, liked it, and it moved fast into vendor selection. Sarvam later handled it.
why I left
By the time I left, the idea I had started was running across two states, had a central adoption, and had three startups grown out of it. It had its own team. It had gotten big enough that someone more senior needed to carry it.
I kept thinking about something I heard about digital public infrastructure: you plant a seed, you grow the tree, you let the forest grow, and then you go find the next problem. That is what I wanted. I wanted to see the supply side, how startups actually build. So I left, and handed it on.