About

Financial rigour tells you what to build.
Engineering tells you how.

I am a Chartered Accountant and CFA charterholder who spent years inside institutional finance and now builds AI tools for the problems that finance left unsolved — mostly Indian ones, mostly boring, mostly the sort nobody builds because the market looks too specific.

The arc

  1. Institutional finance

    Senior analyst, global investment bank

    Research and analysis at the level where being wrong is expensive and someone always checks. This is where the habits came from: show the workings, distrust a number you did not build, and assume the interesting thing is in the footnote.

  2. Markets

    OTC

    Over-the-counter — instruments negotiated between two parties rather than bought off a screen. It teaches you that the terms are the product, that documentation is not paperwork, and that the risk usually sits in the clause nobody read closely.

  3. Product

    AI product manager

    The translation seat: between what a model can actually do and what a business believes it can do. Most of the job is refusing to call the demo finished — the thing that works once, in front of an audience, on a clean input.

  4. 2023

    The first automations

    Started on n8n, wiring boring things together because I was tired of doing them by hand. No grand plan. It was the year the gap between "I have an idea" and "the idea runs" got small enough to cross alone.

  5. Now

    Building in public

    Agentic systems, mostly with Claude Code. One build at a time, including the ones that break, and most of it open source so the claim can be checked rather than believed.

Double entry

Every debit has a credit. The two halves of this are not separate careers — each thing finance taught me turned out to be the thing that made the code work. Read across.

Materiality — knowing which of forty numbers actually matters
Scoping — a tool that does one thing instead of forty badly
Reconciliation — two sources must agree, or you find out why
Evaluation — model output checked against ground truth, every run
Professional scepticism — assume the number is wrong until it survives
Refusing confident output — the single most useful reflex with an LLM
Documentation — if it is not written down it did not happen
Specs and receipts — the screenshot, the number, what broke
Knowing what the answer should roughly be
Catching a plausible, well-formatted, completely wrong result

On paper

  • CA Chartered Accountant — India's hardest professional exam. Audit, financial law, taxation.
  • CFA Chartered Financial Analyst — valuation, portfolio theory, equity research.
  • Azure AI Microsoft Certified: Azure AI Engineer Associate — ML services and production AI pipelines.

Why there is no name here

You will not find my name, my employer or my face on this site. That is deliberate and it is not mysterious.

I would rather the work be checked than the CV be trusted. A credential invites you to believe a claim; a repository lets you verify one. Almost everything here is open source precisely so that the interesting question is "does it run" rather than "who is he."

There is a second reason, and it is practical: finance is a regulated profession, and separating what I build from where I have worked keeps both clean.

How I work

  1. Receipts, not takes

    A take is a sentence; a receipt is a screenshot. If I cannot show the thing running, the number I measured, or what broke, it does not get published.

  2. Build the boring half

    The demo is thirty seconds. The work is the long tail of formats and edge cases behind it. A tool that only handles the clean input is not finished, it is a screenshot.

  3. Read and flag, never advise

    Anything touching markets surfaces what a professional would look at and stops there. It does not recommend, rank or transact — that line is structural, not a disclaimer.

  4. Build for the context I actually live in

    Indian bank formats, Indian tax law, Indian labels, Indian hiring. Global tools treat these as edge cases. Here they are the whole case.

  5. Publish what broke

    The failures are the part almost nobody writes up, which is exactly why they are the part worth reading.

Work with me

GenAI product design, agentic systems, fintech AI.

A small number of consulting engagements at a time. If the problem is interesting, say so plainly — I read everything.

ca.who.codes@proton.me ca.who.codes@gmail.com press & collaborations