Magic That Survives Monday Morning
On Wednesday, the Fed hiked rates for the first time since 2023. On Friday, we typed one sentence into Sortefi:
“Show me US stocks up 2%+ since 2pm ET Wednesday, still above their VWAP anchored to the Fed announcement, on 2x volume.”
40 seconds later: seventeen tickers, ranked. An AI read the sentence, wrote a program to compute an indicator most screeners don't even offer, ran it across thousands of symbols against live data, and answered.
That's the magic trick.
Here's the uncomfortable part: the trick is no longer impressive.
Pretty is free now
Anyone can wire an LLM to a data feed and ship that demo in a weekend. It'll look incredible. Once.
AI made impressive free. It did not make right, every time, with receipts free. If anything, it made it rarer.
A trader doesn't need magic once. They need the same answer at 9:14 tomorrow morning. And the morning after. In fintech, reliability is the product. Everything else is a launch video.
What our system did about itself
The wild part of Friday wasn't the screen. It was this:
Our platform audits every answer before you see it — it reads the AI-written code and checks it did exactly what you asked. On this query, the auditor raised a doubt about the VWAP condition.
The doubt was wrong — the code was right. So within one hour, two fixes shipped: the auditor got smarter, and anchored VWAP became a permanent, exchange-exact primitive that every future query uses.
We ran the same sentence again. Cleaner answer. One pinned convention. Forever.
A user's query at 1pm made the platform more reliable by 2pm. Nobody had to notice. That's the job.
The invisible product
Behind every 40-second answer:
- —Every response audited against your actual request. A deterministic check reads the generated code and verifies every condition you asked for left a footprint — before you see the answer.
- —Data lag? Disclosed. Wrong-market symbols? Quarantined.. If intraday data trails the wall clock, the result says so. If equity-linked perps get caught in a crypto screen, they're separated out with a note to screen them in their native market.
- —An overnight AI judge re-examines every real query. Its only job: did the code actually do what the human meant? Failures get clustered, ranked, and fixed.
- —Suspiciously broad matches trigger a warning. A screen matching 90% of the market usually means your filter is looser than you think. We tell you.
None of this demos well. All of it compounds. Every query makes the next one sharper — and that is the thing a weekend demo can't copy.
The bigger bet
Screening is the wedge, not the ceiling. Plain language in → verified computation over live market data out generalizes to all of fintech: analytics, risk, event studies, research that used to take a quant team an afternoon — done in one sentence.
But it only generalizes if the trust does. The models are rented. The code generation is commodity. The moat is the boring layer: verification, and the honesty around it.
Magic that works once is a demo.
Magic that survives Monday morning is a product.
We're building the second kind.