←  Back to portfolio
AI Trading Copilot · how it works

A calculator that explains itself.

Seven pictures, no jargon. If you only read one section, read the first one — everything else follows from it.

01
The one rule

The AI is not allowed to do maths.

Every number you see — a score, a price, a level to watch — is worked out by rules I wrote. The AI only writes the sentences around those numbers. It can't invent a figure, because it's never asked to produce one.

no way back THE NUMBERS SIDE Rules I wrote plain arithmetic, no model involved 86 ₹1,774 ₹1,707 numbers go across THE WORDS SIDE The AI writes sentences, never figures A checker every figure compared to the record What you see on screen numbers from the left · words from the right · nothing unchecked

The wall only opens one way. Numbers can travel to the AI so it knows what to write about. Nothing the AI produces ever becomes a number.

02
Every night

While you're asleep, four things happen.

Nothing is worked out while you wait. By the time you open the app in the morning, all of this has already run.

STEP 1

Collect

Prices, trading volumes and the day's news for 233 stocks.

STEP 2

Score

Four checks per stock, worked out by rules — never by the AI.

STEP 3

Explain

The AI turns those scores into sentences a person can read.

STEP 4

Check

Every figure in those sentences is compared back to the record.

03
The score

Four questions, asked about every stock.

The score out of 100 isn't a black box. It's four straightforward questions, each answered by a rule, then added together.

Which way is it trending?
Where the price sits against its own recent history.
Technical
How fast is it moving?
Whether the move is picking up speed or running out of it.
Momentum
Is anyone actually buying?
Real volume and delivery, not just quotes on a screen.
Volume & delivery
How rough is the ride?
How violently it swings, so a good score isn't a reckless one.
Risk
04
When something doesn't match

You never see a wrong number. You see duller writing.

If a figure in the AI's sentence doesn't match the record it came from, three things happen — in this order, automatically.

1 · Compare each figure against the record matches? it ships if not 2 · Ask again one more attempt, then stop matches? it ships still not 3 · Plain wording pre-written text, no AI at all always correct, just plainer And when the whole explanation can't run, the app says so on screen instead of pretending it did.
05
How it improves

It marks its own homework. A person grades it.

Every signal is written down when it fires. Later, once enough time has passed to know what actually happened, the two get compared. If that suggests the rules should change, the system can only suggest it.

A signal fires written down instantly Time passes days, not minutes What really happened the honest answer Compare the two was the signal any good? THE ONLY WAY IN A person decides nothing changes itself approved changes only

The system is allowed to notice that a rule isn't working. It is not allowed to change the rule. That decision has a name attached to it every single time.

06
What it costs to run

One dollar sixty a day.

Scoring 233 stocks, writing every explanation and checking every figure — all of it, for less than a cup of coffee.

$1.60 / day

It started at $6.40. I found the cost by tagging every single request rather than reading one bill at the end of the month — and most of the money turned out to be going on work nobody had asked for.

$6.40 → $1.60  ·  three quarters removed, and the writing got better
07
What it can't do yet

The honest list.

A page like this that only describes strengths isn't an explanation. It's an advert.

Some signals are barely better than a coin toss right nowAnd the app prints that on the screen next to them, with the number of past cases it's based on. That's the point of measuring.
Nobody has tested it with a crowdIt runs comfortably for a handful of people. It has never been asked to serve a thousand at once.
Proper back-testing isn't possible yetEarly on I recorded outcomes after the fact rather than at the moment each signal fired, so results can't be tied cleanly to the version that produced them. Fixing that comes before anything else.
Nobody pays for it — deliberatelyIt stays free until a longer run of live results says it's worth charging for.
←  Back to portfolio The full case study  →