What three years of trading taught me about building AI products

Before I built AI products for banks, insurers and consumer platforms, I spent more than three years as a quantitative researcher at a trading firm in Gurugram. I built models that forecast volatility, read options flow and estimated the odds of index moves, and I watched them meet real money every trading day, thousands of lots at a time.

Trading is an unforgiving teacher. A model that looks brilliant in a backtest can lose money within a week, and the market tells you so immediately. The habits I picked up there turned out to be the best preparation for building AI products, where the same mistakes are easier to hide and slower to surface.

A forecast is a probability

In options, you are paid for pricing the odds better than everyone else, and direction alone earns nothing. My index models never said the market will rise. They said something closer to: a 60 percent chance of a move larger than one percent this week. Every position was sized to that number, and to the 40 percent where we were wrong.

Language models are the same kind of machine. They return the most likely answer, in a voice that sounds equally confident whether the answer is right or wrong. The AI products that work treat every output as a probability: they measure how often it is right, decide in advance what happens when it is wrong, and send the uncertain cases to a person.

Backtests flatter, and so do demos

Every quant learns to distrust a beautiful backtest. With enough parameters you can fit any history, and the smooth curve says more about your effort than about the future. The only defence is data the model has never seen, and the patience to wait for it.

A demo is a backtest on hand-picked examples.

AI demos carry the same flaw. When we review conversations for safety or apps for dark patterns, the number I trust is accuracy on a sample nobody chose: drawn across languages and categories, checked by people, and measured again after every change to a prompt or a model.

Risk first, returns second

The traders who lasted were obsessed with the downside. Before a position went on, we knew what it would cost if the market moved sharply against us. I built an options simulator for exactly that question: what happens to the whole book under each scenario we could imagine.

In an AI product, the downside is a wrong answer that someone acts on: a complaint misread, a customer wrongly flagged, a safety issue missed. The design questions are the same. What is the worst plausible failure? How often will it happen? What limits the damage when it does? Human review, confidence thresholds and decision logs are the stop-losses of AI.

Automate the boring middle

A surprising share of my time went into plumbing: automating the path from pre-trade checks to post-trade analysis, so people spent their attention on judgement instead of reconciliation. It cut manual oversight by more than half, and, more importantly, it made mistakes visible the same day.

The best AI products I have worked on do the same. They take over the repetitive middle of a process, leave the decisions with people, and record every step so that problems surface early.

Slow compounding beats clever trades

Alongside the trading desk, I managed long-term equity portfolios with a CANSLIM-style approach: own leading, high-quality businesses and hold them through the noise. It taught me that steady, repeatable decisions compound, and heroic ones rarely do. That belief is why I later qualified as a SEBI-registered research analyst, and why I care about helping retail investors build wealth patiently.

Markets gave me fast, humbling feedback. AI products deserve the same discipline, applied before a customer applies it for you.

Nothing in this essay is investment advice.