I build rooms. At eighteen it was a chat room of ten thousand people who liked a programming language, the newest one has forty founders in it, and the ones in between were for machines. Every job I have had started as a side question I could not put down, which is a polite way of saying the plan was never the plan.
Read this top down for cause and effect. The project pages are the tidy version, this is the one with the turns left in.
now Independent researcher and product engineer in Hyderabad. Memory a person can audit, in the open at Marshmallow, and systems that have to show their reasons, at Biome. Groups working on agents, memory, and human + AI interaction. Building the future of work and how society shapes around it. Open to opportunities to collaborate.
The asking that never stopped
My father is an engineer, and I was the kid who followed him around asking why engines sound the way they do and why buildings stand the way they stand. He answered most of it, which was his mistake, because a child who gets answers keeps asking. Twenty years later I built him a tool for his interior design business, still my favourite deployment story, a few entries down. The asking is the only constant on this page.
JuliaCon, at eighteen
First year of college, and I had a thing for programming languages nobody around me used. Julia was the one that stuck, so I lived on its Discord and got talking with Viral Shah, who co-created it. That summer JuliaCon went online for the first time and Viral asked me to help moderate it, about ten thousand people in one chat, and I did it the way you do anything at eighteen when someone you admire asks, completely. This is where I learned the thing I keep relearning: a room holds together on a handful of norms and a handful of people who care enough to enforce them kindly. I have been building rooms since. The machines came later.
From scratch, repeatedly
I rebuild things to find out what they are. A 2.5D Doom in bare C++ in first year, the walls drawn in ASCII because I had no graphics library and refused to get one. A Linux system compiled package by package one bored summer, then a window manager bent to my habits and a dmenu rewritten to index files by tags and contents; the slang for the visible part is ricing, and I was in far deeper than the wallpaper. Ethereum's 2016 whitepaper reimplemented in Go. IBM's quantum summer school, taken because it looked impossible, and impossible mostly turned out to mean unread. Later, PyTorch's autograd rewritten to see the machinery. None of this was for anyone. It is the same instinct that made me write an operating system for agents instead of picking a framework, and it keeps landing me somewhere I did not plan to be.
Phi Phenomenon
My college did not have a research community, so in second year I started one and named it after an optical illusion I was fond of, two lights blinking in turn that your brain insists is one light moving. Reading groups, paper walkthroughs, people who wanted to build things nobody had assigned. It grew into one of the bigger clubs on campus and I handed it to juniors, which is the whole point of a college society. It drifted from what I meant it to be after a couple of handovers, and it is still running, still the only research society the college has. A thing that outlives your attention is a different kind of success from a thing that grows, and I did not know that yet.
Venture, from the inside
I wanted to know how money thinks. About ten months of fellowships across venture firms in Southeast Asia and Rbl1 in New York, reading reports, sitting with valuations, pulling industries apart to see how they were put together. At Rbl1 I was put with two other fellows and ended up leading us, because I was the one who could not shut up about the new blockchain industry, and the thesis we wrote made it into a few company memos. I went in wanting to know how money thinks and came out knowing where businesses actually break, the more useful thing. It also left a question that took four years to answer: what would I build if I ever sat on that side of the table.
The year I lost
Third year, straight out of the lockdowns, and I was out, eating anything from anywhere and sleeping when I fell over. I got sick enough that I could not sit exams two semesters in a row, and my transcript still shows it. That was on me, nobody else. What I did about it was cycle every day, cut carbs to almost nothing, and go from 130 kilograms to the low nineties, forty kilos gone and kept off. The lesson was embarrassingly simple, a body is a computer and you get out what you put in, and the grades came back once the input did. I would rather have learned it cheaper. It is still the most useful year on this page.
Circles
It started as a group chat where I posted internship openings for anyone who wanted them. People kept joining, so I built it a home, Circles, the network where campus events and parties got hosted, with a five percent commission on the way through. About 5,000 of a twenty-thousand-person campus, real revenue, and a plan to be the Reddit of Indian college networks. Then I got comfortable. The growth was nice, the parties were nice, and I stopped thinking about what came next, which is the failure I own here. A mentor told me plainly that I needed to learn a lot more before I was the best at anything, and when Slick, which had just announced a $1.6 million seed, offered to fold Circles in, I took it. A small fish in that pond beats the biggest fish in a pond of one campus. fig. p-07.
A language model inside a browser tab
Final semester and an internship at Tmotions, where the most interesting problem in the building was local inference inside Brave. I compiled ggml and llama.cpp to WebAssembly as internal R&D, wrote the JavaScript adapter that pretended to be an operating system, and watched tokens come out of a browser tab with no server behind it. It worked. It also forgot everything the moment you closed the tab, and I remember being more bothered by the forgetting than pleased by the tokens. That is the problem I am still on. fig. p-08.
Slick
Slick acqui-hired Circles, and me with it, which is a funny way to get your first product job. Archit, the founder, gave an intern with context from both sides of the deal the run of the product surface, and after six months I was the one person on product full time while the network went from zero to 2.5 million users, the company's announced figure. Strategy was his call, the surface was mine. The cold start got solved with presence the app generated itself, events, nudges, rewards, games, a few clearly automated update accounts, holding the room until real people arrived, and at night I taught myself the machine learning I was shipping by day. Then the founders read the numbers, Instagram absorbing everything and nobody paying for social, and rebuilt the company around quick commerce, the right call for them in hindsight. I did not want to build that and it had no AI in it, so we parted well. Those were the sporadic startup chasing years, and they left me the question I still think about most: a feed that is partly synthetic should say so inside the product, before anyone thinks to ask. fig. p-06.
HCLTech, the enterprise chapter
I moved back home and joined HCLTech's new GenAI division to see what building looks like inside a company of two hundred thousand people. Lead engineer on an LLM guardrailing, evaluation, and red-teaming platform that shipped to Fortune 500 clients, three engineers to train and assign for the first time, and a lot of pitching, which nobody tells you is most of the job. My first version scored answers with a number nobody could explain, and a buyer cannot defend a number nobody can explain, so the score got rebuilt into things a person could argue with: truthfulness against a corpus, relevance to the question, a policy screen. Then, in 2025, someone close to me died. It rearranged what I thought I had time for. I handed in my notice in May and spent the three months of it, and most nights before that, on Five Labs. fig. p-05.
Five Labs
The work needed a home, so in 2024 I built one alongside the day job and wrote it a manifesto about acceleration and open access. Small tools for everyday people first, kept free on purpose, and InteriorBot for my father's business, the deployment story I promised you at the top. Then an open-source Operator clone, a framework for giving any model emotion and memory, a benchmark for social reasoning, and a picture of agent swarms coordinating like honeybees took over the year: months of reading, a Berkeley agents course, a Discord group chasing the platonic representation hypothesis with nothing published and everything learned, and a whitepaper called Indra, an operating system for how a swarm should reason, with a public MVP beside it. From May to August 2025 it was the only job I had. The unit economics did not hold and I shut the studio down; the paper and the code are the record. The manifesto had been about the wrong thing. Two years on, the essay I wrote was about trust. fig. p-02.
Biome
Zaran Bhagwagar found me, and that is the door. He brought me into Biome, a venture fund, as an entrepreneur in residence to see whether the Five Labs thesis could hold as a business, and was my thinking partner the whole way through. It could not; that verdict is mine, not the fund's. So I moved into the product team as its only technical person, answering to Srinivas Kollipara, and the reasoning work got a different body: a sourcing platform, an evaluation agent, the databases and knowledge system the fund actually runs on. The first ranking I shipped got ignored, because nobody could see the reasons for it. The one that gets used shows every check it ran. That sentence is the whole site. And four years after the fellowships, I was sitting on that side of the table. fig. p-03.
Intersection
Late in 2025 I started the room I had wanted to be in. Intersection is a closed group of forty founders under twenty-five, and the range is the point: robots for manufacturing in low earth orbit, a next-generation implant company, a two-person lab training small models for Indic languages that holds its own against far bigger names. Between them they have raised about ten million dollars, by their own count, for their companies, not for the room. I run it the way I moderated JuliaCon, a few norms, kindly enforced. It is where I check whether anything I believe about trust and machines survives contact with people who are actually building.
The essay
In February I wrote the Pixarification of Life, and the question finally had words. We are about to put life into objects and coworkers that never had it, and the people it is arriving for do not trust it, for good reasons. Trust like that is not argued into someone. It is earned across repeated contact by a thing that remembers you and shows its work, memory and legibility, the two things every entry above kept tripping over.
Marshmallow
The question Indra left me with was smaller than a colony: what does one agent need to know before it walks into a room with you. Marshmallow is the answer I am building in the open, plain files, source-backed, nothing learned without your say-so, byte-exact rollback when it learns wrong. It runs under Claude Code, Codex, and Cursor, and this page was written with it watching. I had been circling it since a language model sat in a browser tab with nowhere to put what it knew. fig. p-01.
Larch Hollow, one night
A world-model hackathon and a murder mystery you solve on foot. The model got to invent the weather, the app kept the crime still, twelve hours, and the clearest thing I learned all year about where generated worlds need a ledger. A room where the storyteller is a machine has the same problem as a feed that is partly synthetic: somebody has to hold the record. fig. p-04.