The engine
Written for someone who wants to understand it before trusting it.
People ask two things: what does it actually do all day, and what stops it doing something stupid. This page answers both properly.
Exact thresholds, multiples and caps are not published. Not because they are magic — they aren't — but because publishing a live system's precise trigger points invites others to trade against them.
The method below is not a secret, and most of it is textbook. The product is the discipline of applying it without flinching, and the machinery that catches it when something breaks. If someone tells you their edge is a clever formula, be sceptical.
Once per scan the desk narrows roughly five hundred NSE stocks down to a handful worth looking at, asks a language model to pick at most one of them or decline, sizes that position from its stop rather than from a fixed rupee amount, places the order at a price that will actually fill, then manages it to a stop, a target or a time limit. Every few minutes it compares its own record against the broker's, and a second, separately written program checks both.
On most days it finds nothing and does nothing. That is the intended behaviour, not a fault.
Before any model is consulted, plain arithmetic removes almost everything. A stock has to clear all of:
This stage costs nothing and typically leaves a handful of names, often zero. When it leaves zero, the desk stops there and no model is called.
Survivors are enriched with the things that change what a breakout means:
Only now is a language model asked, and only about the shortlist that survived. It is given the candidates, the market context, and the desk's own written lessons from past trades. It returns one pick, or NO_TRADE.
This is the part people are rightly suspicious of, so precisely what it can and cannot do:
Pick one name from the shortlist, or refuse them all. Declining is common and entirely acceptable.
It cannot set or widen your stop, change your position size, raise a capital cap, or override a concentration limit. Those are computed outside it and it is never shown a path to them.
It may only name a candidate from the list it was given. Anything else is rejected by schema validation before it reaches the order path.
If it is throttled, over budget or errors, the result is no trade — never a fallback entry on rules alone.
That last point is deliberate and was a correction. An earlier design let the deterministic rules enter by themselves when the model was unavailable — which meant the desk traded hardest exactly when it knew least. It no longer does that.
Model spend is governed by a minimum interval between calls and a daily token budget, per strategy. It runs on efficient open models, and the cost is a few rupees a month per account.
This is the most important paragraph on the page.
Most people decide how much to buy, then wonder where to put a stop. The desk does the reverse. The stop is placed first, derived from the stock's own daily range — wide enough that ordinary noise won't hit it, tight enough that being wrong is survivable — and then bounded by a hard ceiling so a wild stock cannot justify an absurd stop.
Only then is the quantity computed, as whatever makes the distance to that stop equal to a fixed slice of your capital. The consequence is that the rupee risk per trade is roughly constant whether the stock is calm or violent. A volatile name simply gets a smaller position.
The target is set as a multiple of that same stop distance, so the reward is always expressed relative to what is being risked rather than as an arbitrary percentage.
Orders are priced at the live touch — the bid or the ask as appropriate — rather than the last traded price. This sounds pedantic and is not: pricing at the last trade means an order only fills by luck, and an exit that doesn't fill is not an exit.
The desk waits for a real fill confirmation rather than assuming one, and a partially filled order is recorded as partially filled. On an exit that isn't filling, it will step the price through the touch by a small increment rather than sit there — getting out matters more than a few paise.
Implausible price prints are rejected before they can be acted on. That guard exists because a bad print once reached a position's valuation and briefly made the book read catastrophically wrong.
Software that trades your money should assume it is wrong. Every few minutes:
The second reader exists because of a real failure. A bug in the desk's own reconciliation once invented a holding that did not exist, and it survived for days — because there was only one implementation of "what does the broker hold", and every part of the system read that same wrong answer. One implementation cannot check itself.
The desk includes a transformer model trained on price data, which produces a next-session forecast for the names on the shortlist.
Tested offline against a year of our own NSE history across thousands of walk-forward windows, its rank correlation with actual next-day returns was small and positive, and it beat a naive momentum baseline that was itself slightly negative on the same data.
"Small and positive" is the accurate description and we are not going to dress it up. It is treated as one input among several that can tip a close call, never as a prediction to act on. Its accuracy in live use is measured continuously and shown alongside the offline figure, so if it stops working we find out from data rather than from a bad quarter. Removing it is a one-line change and no risk rule depends on it.
Being straight about this matters more than the rest of the page.
The signal is not the edge. Volume-breakout momentum is well known, widely traded, and works in some regimes and not others. Expect losing runs. Nothing here has an information advantage over the market.
The edge, such as it is, is behavioural. The stop is set before the position exists and is not moved wider because the trade is going badly. The size is computed rather than felt. The loser is cut on a schedule. It does not get bored on a quiet day, or brave after a winning one. Most retail losses come from doing exactly those things — and a program is simply better at not doing them than a person is.
That is a modest claim. We would rather make a modest claim you can check than a grand one you cannot.
Still have a question? The FAQ covers more, and you can ask directly — including the awkward ones.