CASE STUDY 01 · AI × FINANCE · DESIGN + BUILD
Making AI-written SQL something an analyst can actually sign off on.
00 — THE LIVE APP
01 — THE PROBLEM
01
A language model can write SQL in seconds. The hard part is trust: nobody can sign off on a number when the assumptions behind it are buried in a paragraph of explanation.
Nexus treats AI output like work pinned up for a design critique. Every assumption is marked up where it applies, every change becomes a new version, and nothing ships without a human sign-off.
02 — THE BUILDER, MARKED UP
Which non-bank companies saw gross margin fall three years running — and did insiders buy anyway?
Not a summary paragraph under the result. A reviewer can reject one premise without rereading the whole query.
What the model says about itself sits beside independent system checks. When they disagree, that's the most useful thing on the screen.
Accept, edit, reject are the only ways a transformation moves forward. Every version is kept and can be diffed.
Requested, generated, repaired, edited, accepted — AI actions labeled separately from human ones, exportable as CSV.
03 — MEDIUM & SOURCES
| AREA | SOURCE | REFRESH |
|---|---|---|
| Fundamentals | SEC EDGAR XBRL frames | WEEKLY |
| Insider trading | SEC Form 4 data sets | DAILY CHECK |
| Fund holdings | SEC Form 13F data sets | DAILY CHECK |
| Prices & valuation | Finnhub | DAILY |
| Macro | FRED | WEEKLY |
| FX | U.S. Treasury | WEEKLY |
04 — KNOWN LIMITS
Being upfront about the limits is part of the design. These are shown to users, not hidden in a README.
05 — OUTCOME
It started as a six-week concept — the AI Transformation Builder. This is that concept made real: a working app running on live public SEC data, with every decision from the original design intact.
06 — NEXT ITERATION