Lokesh Basu

Hi, I'm Lokesh. I live in Bengaluru with my wife Riya, our kids Gaurik and Vaani, and Oreo, a St Bernard. I've been writing code since school, and these days I build products that use language models.

GitHub · LinkedIn · X · lokesh.basu@gmail.com

How I got here

I wrote my first programs in BASIC in a school computer lab, around 2000. Lots of 10 PRINT and GOTO. Then Java, then C++, then whatever the next thing needed.

I studied computer science at IIT Roorkee from 2010 to 2014, and spent more time building than studying. I set up the department's Hadoop cluster, made a health search app for doctors and clinics, and came second at a Microsoft hackathon with a game we built in 24 hours.

After college I worked on the Android rendering framework for Gear VR at Samsung, co-founded a small healthcare search startup, and was an early engineer at SmartVizX on Trezi, a multi-user VR design tool.

Then came a few years of data and ads: insurance data at Rippling, the ads data system at Disney+ Hotstar (300 million users), and display ads at Udaan.

From 2022 I led teams. At Hike it was payments, wallet and fraud, with about a million dollars a day moving through our systems. At Amazon I led 16 engineers on the platform behind Amazon Now, the 10-minute delivery service. Both jobs taught me the same thing: the hard part of any automated decision is noticing when it's wrong.

In 2026 I left to build on my own. Most of my work now is plain engineering around language models: limits they can't override, checks on what they say, and knowing what each call costs.

What I'm working on

Proofsource: measures what ChatGPT, Claude, Perplexity and other AI engines say about a brand. It shows error bars, and says "not enough data" when it doesn't know.
Pagerless: on-call for apps. Agents try the fix first, and a person gets paged only when one is really needed.
Rivane: an accounting OS on top of ERPs, which I co-founded. The code does the maths; the model only explains it.
Driffle: meeting notes.
g4a: keeps an AI coding agent's reasoning in git, so you can see later why the code changed.

One thing from this year I enjoyed: I fine-tuned small open models on my laptop and tested them against Claude Sonnet 5 on three classification tasks. The small models won one and tied two, at 30 to 51 times lower cost, and the whole study cost $7.83 in API calls. Everything is on GitHub, including the ties.

Family

Riya and I have known each other since school. Gaurik is our son, and judging by the photos, his interests are Pokémon, colouring and rolling around on beaches. Vaani is our daughter and the youngest of us. Oreo outweighs both kids put together and decides where most of our walks go.

Gaurik and Vaani smiling from the car windows, with Oreo the St Bernard in the front seat — Gaurik, Vaani and Oreo, ready to go somewhere.

Riya smiling and holding a takeaway box, with Gaurik in a Pokémon t-shirt beside her — Riya and Gaurik, on snack duty.

Gaurik walking Oreo on a leash, seen from behind — Gaurik walking Oreo. Or the other way round.

Oreo greeting another dog on a morning walk — Oreo making a friend.

Gaurik's colouring of a Channapatna wooden horse toy — A Channapatna horse, colouring in progress.

Gaurik doing a forward roll in the sand — Beach day.

Say hello

Email is best: lokesh.basu@gmail.com. I'm always happy to hear from people building with AI, or anyone who wants to talk about getting these models to behave.