About

I do not design from the brief.

Every system I have built began the same way. Somebody hands me a description of the work, I go and look at the work itself, and the two do not match.

A compliance team told me their problem was speed. Thirty to fifty minutes to compare two near identical data files against a two hundred page rulebook, and they wanted it faster. I spent time with the analysts before designing anything, and what I saw changed the project twice over.

They never reasoned about the customer’s name or address. They reasoned about the structure around it. So the sensitive data was not an obstacle to putting an agent on this, it was something the agent never needed to see.

And they did not want a faster answer. They wanted one they did not have to check twice. An answer you cannot verify just moves the work somewhere else.

Neither of those was in the brief. Both of them are now the architecture.

How I work with clients

I am independent and I work asynchronously by default. That is not a preference I am asking you to accommodate, it is how the delivery actually runs.

Every project I run keeps a written spine: self contained briefs that assume no shared context, locked data contracts, a written definition of done, and an append only ledger of numbered rulings. The mapping system has 205 of them. Nothing important lives in a conversation, because context that only exists in a conversation is context that will be lost.

There is a standing rule underneath all of it: when the measurement disagrees with the plan, the measurement is right. That has overruled my own briefs repeatedly. One of them asserted nine unresolvable types. The reachability walk proved eight. The record says eight.

I built that discipline because agents have no memory across sessions and a strong bias toward declaring success. It turns out to be the same discipline a distributed team needs.

I work from Hyderabad, across APAC and EMEA hours.

What I am hired to decide

The decisions are narrower and more useful than “AI”:

Where the model sits, and what it can reach. In one system it is invoked exactly once, inside a redaction zone, after a human has signed off on the exact payload. Three components in another are barred from calling a model at all, permanently, by rule.

What happens when it is wrong, and how the person finds out.

Where the human belongs. Checkpoints go where a mistake is still cheap to fix, not where they feel reassuring.

How you would prove any of it afterwards. Per call ledgers, prompt fingerprints, every override logged.

Whether the measurement can be trusted at all. More on that below.

How the code gets written

AI writes it. I direct, review and own every architectural, product and evaluation decision, and my shipped standard is that I wrote, modified or explicitly reviewed every line that shipped.

The mapping system is 22,701 lines of Python with 947 passing tests, built in 26 days, and I never opened an editor.

I am still ramping on Python as a hands on skill and I am not going to pretend otherwise. What I bring is the architecture, the evaluation design, the product judgment and the client facing work, plus a delivery method that has produced three shipped systems.

The one thing I build against

Silent failure.

An agent that visibly breaks is a nuisance. An agent that drops something and carries on as though nothing happened is dangerous, because nobody finds out until a customer does.

A golden set audit told me one of my pipelines was covering about thirty percent of a lesson and quietly discarding the rest. Nothing had flagged it. On another project I built an evaluation harness, then found it could settle only 2.5% of the real disagreements between two candidate models, which meant it was measuring agreement with one analyst rather than correctness. I changed the product because of that number.

Those are the two findings I am proudest of, and both of them are embarrassing. An audit that humiliates your own pipeline is an asset, not a setback.

The network on the front of this site is not decoration. A colony is a real multi agent system: no central controller, coordination through traces left behind, and every so often a worker drops its load and keeps walking while nobody notices. That is the failure I design against, drawn.

Before this

  1. Independent · April 2026 to now

    Full time on client builds.

  2. Founder’s office type role · Meril

    One of India’s largest healthcare and medical device companies, and the role was a founder’s office one rather than conventional product design. I implemented AI tooling across departments, ran internal AI education, and moved the organisation off manual data entry. That job is where I learned how a large company actually adopts a new tool, which is a completely different problem from building one, and it is the reason I talk about change management as part of shipping an agent rather than as somebody else’s job.

  3. Design project lead · Svasara, Tadoba tiger reserve

    A lodge inside the reserve. Different industry, and I am not going to pretend it was AI work. What it did give me was a year of designing for an operation I had to observe closely before I understood it at all.

  4. Design intern · Favo Robotics

  5. Human Computer Interaction · IIT Delhi

  6. Bachelor of Design, Transdisciplinary Design · Anant National University

    Which is a long way of saying I was trained on problems that refuse to sit inside one discipline.

I write occasionally. The piece I would point you at is Stop Confusing “Using AI” with “Designing AI”.

Away from the desk

I ride. Off road when I can find it, and I am travelling around India at the moment, which mostly means arriving somewhere with no plan and asking people where to eat.

I am a wildlife person. Safaris when I can, and slow rides through forests for no reason other than being in them. The year I spent working inside a tiger reserve was not an accident, it was me arranging my life around the thing I like most.

What actually drives me is novelty. New places, new food, new processes, new tools, anything at the edge of what I currently understand. It is why I ended up here rather than in a settled career, and it is why I like the microscopic end of nature best, the parts operating below the threshold where anyone is paying attention. That preference turns out to describe my professional obsession too, which I did not plan.

One project is genuinely personal. I wanted proper visual educational content when I was a kid and it did not exist for me. I build it now, pro bono, for one school, for children between four and ten. It is the only thing I work on where the audience cannot tell me what they need, so I have to be far more careful about what the system is allowed to get wrong.

I speak English, Hindi, Marathi and Telugu.