Anupam Choudhari

About

Last updated 15 October 2026

Anupam builds AI products. Fifteen-plus years of it, twice as a co-founder, and most of that time spent on the unglamorous half of the problem: taking AI from demos and PoCs to something enterprise-ready, and finding where it creates real value once it gets there. A demo only has to work once. An enterprise product has to work on the hundredth messy request, inside someone else’s security review, for people who never asked for AI in the first place.

These days the work is agent harnesses. Not the model, not the prompt, the scaffolding around them that decides what the agent can see, what it is allowed to do, when it should stop and ask, and how anyone knows whether any of it worked. The interesting problems in AI products have moved from capability to permission, evaluation and trust, and almost none of that is visible in a demo.

Lately, as the models have become commoditised, the harness has started being called the moat, and for good reason: it is where the value gets created. The models are rented and get better on someone else’s schedule. What wraps them, the context, the tool access, the guardrails and the evals, is what fits the work and what the customer can’t swap out. The thing that used to be dismissed as “just a wrapper” has become the thing customers pay for.

Timeline

  1. 2014
    A graph of skills, occupations and traits. A knowledge graph that says who can do what, and what it would take to get from one role to the next. It started as a data problem and turned into a decade-long thread.
  2. 2016
    A workforce mobility platform built on top of it. Internal mobility, from both sides: career pathing for employees, and workforce planning and talent transformation for the employer.
  3. 2021
    A recruitment marketplace, co-founded and scaled. SaaS-enabled, connecting companies with recruiters and boutique agencies: 1,500+ hires for 200+ companies, 5,000+ recruiters on the platform, and more than $750K in revenue for recruitment partners.
  4. 2024
    A generative AI product and engineering studio, co-founded. Agents that got work done from inside Slack, before that was commonplace.
  5. 2025
    A benefits concierge. Used by employees of 2,000+ companies, deflecting 100,000+ queries that would otherwise have landed with HR.
  6. 2026
    Agent harnesses for enterprise workflows. The current work: the layer that makes agents dependable enough for an organisation to actually hand them a job.

Operating principles

  • Stay in the problem. Is it worth solving, and will it still matter five years from now? Avoid the build trap. Write a thesis, be specific about goals, measure with real business metrics, and set milestones with launch dates.
  • Check what’s already solved. Most problems are partly solved somewhere. Reuse the parts that work, and don’t reinvent the wheel, least of all when the homemade version would be worse than the frontier.
  • Show the thing. A working prototype that is wrong is more useful than a document that is right.
  • Get as close to the customer as possible. Every signal counts, but they are not equal: a direct conversation beats explicit feedback through customer success, which beats product usage metrics, which beats simulated agentic usage loops, which beats building on gut. A usage dashboard says people stopped at step three; a ten-minute call says why. The sooner the work moves up that chain, the better the product gets.
  • Build for the model six months to a year out. Capability moves faster than roadmaps. Being early and failing beats being late and coming a distant second.
  • Default to general models. The AGI age is already here, and most applications don’t need a verticalized small model. SLMs built a year ago are already falling behind general models on benchmarks, and replacing them gets expensive, mostly because of the sunk cost.
  • Write the evals before the code. Evals as the PRD is easier said than done; do it anyway. A clear definition of done is what lets an engineering team tell its agents when to stop. Without one, even the best engineers end up back in 2025, hand-coding what an agent could have done, and the whole team slows down. Good evals are what free engineers up for the parts only they can do.

How it all started

A few years in the United States, first for a master’s in computer science at RIT in Rochester. The thesis was on partially connected neural networks, back before anyone had worked out that you could train neural nets on GPUs. A course on Biologically Inspired Intelligent Systems settled the question of what came next: the future was going to be AI.

Then Los Alamos National Laboratory. Days went to bioinformatics research on H1N1, and to a first taste of high-performance computing: running ML training on a few cores of a supercomputer. Evenings went to pool with postdocs from the physics labs, at the only bar in town, which is where the real intellectual curiosity turned up. In between, an unreasonable number of road trips across the country. There’s an easter egg on this site for those; you’ll find it or you won’t.

Back in India, an edtech startup in Bangalore building Funtoot, an adaptive math tutor, is where the love for NLP, startups and building AI products took hold. A tutor that has to work out why a ten-year-old got a fraction wrong is a surprisingly good education in language, data and product all at once.

Home has been Bangalore ever since, after starting out in Mumbai. Bangalore traffic, and where the taxes actually go, get their own section: Mildly Annoying, complete with armchair solutions.

Worth a cheat day or two

  • Pizza with a personality of its own. Fresh ingredients, a real sourdough base, something that tastes like a decision was made.
  • A proper breakfast burrito. Enough time in New Mexico to know what a good one tastes like. Still searching for one in Bangalore.
  • A burger that’s about the meat. Plenty of places in the city have mastered the craft, but in the rush to gourmet everything with freshly baked buns, some forget the patty is the point. Fries are not optional. And not wedges.
  • A good cheesesteak. Rare, and worth the trip when it turns up.
  • Light roasts and IPAs. Coffee has its own page because it got out of hand. Beer may join it.

No seafood, which people treat as a character flaw. Peace has been made with it.

Elsewhere

On social media, strictly a consumer. Posting may come back eventually. The “going direct” trend, with leaders publishing on their own terms instead of through intermediaries, is a genuinely good development, and this site is a first step in that direction. Maybe someday, fully direct.

Say hello. Email anupam dot choudhari at gmail dot com, or book 30 minutes.