A clinical note becomes a structured case with a verbatim source quote behind every fact. A physician reviews the AI draft, peers play it stage by stage, and the server scores every decision, opens the debrief and shows where the cohort turned.
Rare diseases take years to diagnose because the first explanation usually fits well enough. The featured case is built to test the second look.
FastAPI · PostgreSQL 18 · Next.js 16 · Gemini and Claude on Vertex AI · Cloud Run
The case most physicians miss
32-year-old woman · hard
Sudden left-sided chest pain and breathlessness
Each part runs live. Open a card to see it working.
Cases arrive as JSON, land in a normalized schema through Alembic migrations, and every attempt is scored on the server.
Server Components render the case, a client stepper plays it, and a Server Action submits the answers. Types come from the API's OpenAPI schema.
Text is de-identified, extracted against a JSON schema with a verbatim quote for every fact, checked against the source, then authored into decisions.
Raw note
pasted or sampled
De-identify
names, dates, IDs masked
Extract and ground
every fact quoted, every quote checked
Physician review
flagged items, not the whole case
Play
nine stages, scored on the server
Debrief
percentile, AI markers, sponsor view
Nothing is marked until the case closes. The player never shows right or wrong mid-case, and the answers stay on the server until you submit. Points come only from the diagnosis and the treatment plan.
Built after the three tasks were done and tested, around how a case platform actually runs.
Gemini and Claude play every showcase case blinded, through the same scoring code. They show up as labelled markers on your percentile curve and never enter the cohort.
Studio output is only a draft. A reviewer sees the answer key and a server-computed checklist, checks the few flagged items, and signs off before the case counts.
After you close a case: pick rates at every decision point, the right steps most physicians skipped, and how ordering the key test changed the diagnosis rate.
Gemini and Claude extract the same eight synthetic pulmonology notes; every fact must quote the source. A smoke test, not a benchmark, and the page says why.
Your production model vs. the next one
On this gold set Claude Opus 5.5 is about as accurate (+0.01 macro-F1, within the noise of 8 cases), 20% cheaper per case and 1.9 s faster at the median. Worth a shadow run on your own cases before switching.
Claude Opus 5.5, best of 4 models on the same notes. Full comparison