RadPilot

AI-native radiology reporting MVP: a worklist, a reading room with a DICOM viewer next to a structured report editor and a copilot panel, streaming AI report generation from shorthand findings, and a hybrid copilot (deterministic safety rules + LLM review + guideline calculators). Reports go through a sign/amend lifecycle and export to FHIR R4 and PDF.

Demo project. Synthetic data only. Not for clinical use.

Live: https://radpilot.vercel.app (demo login available on request)

Stack: Next.js 16 (App Router) · TypeScript · Supabase (Postgres, Auth, Storage, RLS) · Vercel AI SDK · Tiptap · Cornerstone3D · shadcn/ui + Tailwind v4 · Vitest · Playwright.

Status

All sessions (0, 0b, Docs, Tracks A-D, Session 5) are done: auth and worklist, templates and streaming AI generation, the copilot/guidelines/lifecycle, the DICOM viewer and FHIR/PDF export, and the reading room that wires them together. See docs/sessions.md for how the work was split, and docs/demo-script.md for a 5-minute walkthrough of the seeded studies.

Getting started (local)

Requires Node 22 and Docker Desktop (running). The Supabase CLI is a dev dependency, installed by npm install.

npm install
cp .env.example .env.local
npm run db:start          # local Supabase: API :54321, Studio :54323. Prints the keys.

Fill .env.local with the values db:start printed (run supabase status to see them again): NEXT_PUBLIC_SUPABASE_PUBLISHABLE_KEY (publishable or anon key) and SUPABASE_SERVICE_ROLE_KEY (secret or service_role key). Then:

npm run db:reset          # apply migrations and supabase/seed.sql
npm run dicom:generate    # write the synthetic phantom DICOM to supabase/dicom/ (gitignored)
npm run dicom:upload      # upload it to the private `dicom` bucket (uses the service role key)
npm run dev               # http://localhost:3000

Log in as the demo radiologist: [email protected] / radpilot-demo (an admin, [email protected], has the same password).

Demo data

supabase/seed.sql creates 2 users and 6 report templates (CT chest, CT head, CT abdomen/pelvis, chest radiograph, US thyroid, mammography), then calls seed_demo_data() for 25 synthetic studies. Study times are relative to now(), so the worklist always looks fresh.

The demo has a shared login, so the studies are restored to this set every night at 07:00 UTC by reset_demo_data() (pg_cron), and on demand from the admin's user menu (Restore demo data). Both live in supabase/migrations/20261002000000_demo_data.sql.

Studies 1 to 10 each demo one feature, starting from an unread study:

#	Exam	Patient	Priority	Demonstrates	Images
1	CTA chest	F 58	STAT	Suspected PE: critical finding	ct-chest-phantom
2	CT chest	M 64	routine	Incidental nodule, 30 pack-years: Fleischner	ct-chest-phantom
3	XR chest	M 71	STAT	Left subclavian line, pneumothorax: critical + laterality	cr-chest-pneumothorax
4	XR chest	F 45	routine	Already signed (final): FHIR and PDF export	cr-chest-normal
5	CT head	M 82	STAT	Fall on anticoagulation, hemorrhage: critical + laterality	ct-head-phantom
6	CT abdomen/pelvis	F 27	urgent	Right lower quadrant pain: laterality + sex checks
7	CT abdomen/pelvis	M 59	routine	Cirrhosis, HCC surveillance: LI-RADS
8	US thyroid	F 39	routine	Palpable left nodule: TI-RADS + laterality
9	Mammography	F 52	routine	Screening: BI-RADS
10	CT abdomen/pelvis	M 68	urgent	Hematuria: sex-mismatch check

Studies 11 to 25 are the same exam types in other states, so a visitor can pick up anywhere in the workflow:

#	Exam	State	Demonstrates
11	CT chest	draft	AI draft pending review: Accept / Regenerate, sign gate
12	XR chest	draft	Pneumothorax: signing blocked until the critical finding is acknowledged
13, 25	CT head	preliminary	Ready to sign
14	CTA chest	final, critical	Signed pulmonary embolism: critical flag on the PDF
15	XR chest	amended	Re-opened after signing: edit and sign again
16, 17, 18	CT abdomen/pelvis, US thyroid, mammography	final	Signed reports to export (TI-RADS, BI-RADS)
22	CT abdomen/pelvis	draft	Left/right conflict between findings and impression
19, 20, 21, 23, 24	CT head (STAT), XR chest, CT chest, US thyroid, mammography	unread	Start from scratch

The phantoms are drawn by scripts/dicom/phantoms.ts (a CTA chest with a right pulmonary artery filling defect and an 8 mm right lower lobe nodule, a head CT with a right subdural hematoma, and normal and pneumothorax chest radiographs) and written by a small dependency-free DICOM Part 10 writer (scripts/dicom/part10-writer.ts).

Full walkthrough (local Supabase, OpenAI, hosted Supabase, Vercel) with a check after each stage: docs/setup.md. npm run verify:setup tells you what's missing.

The reading room

Opening a study from the worklist (/studies/[id]) loads a three-pane reading room (src/app/(app)/studies/[id]/):

• Viewer (left) — Cornerstone3D stack viewer loaded from the seeded DICOM bucket, with window/level presets, zoom/pan, stack scroll and a length tool. "Insert into Findings" appends the measurement to the report's Findings section.
• Editor (middle) — the Tiptap report editor. "Generate draft" streams an AI draft from shorthand findings into the AI-assisted sections; edits autosave (debounced, optimistic concurrency on reports.version).
• Copilot (right) — every edit re-runs the deterministic rules engine (and the LLM reviewer, when wired) after a short debounce; one-click fix, dismiss, and a blocking banner for critical findings.

The header has Sign (runs the sign gate first, with a friendly message if blocked), Mark preliminary, Amend (once final) and, once the report is final or amended, Export FHIR / Export PDF links to the Track D routes.

Panels are resizable (shadcn's resizable, wrapping react-resizable-panels).

Scripts

Script	What it does
npm run check	typecheck + lint + unit tests
npm run test / test:watch	Vitest
npm run test:e2e	Playwright, including the reading-room happy path (e2e/reading-room.spec.ts)
npm run eval	AI eval suite: 18 golden cases scored against the configured LLMProvider (evals/)
npm run db:start / db:stop / db:reset	local Supabase (reset re-applies migrations + seed, the "reset to demo state" script)
npm run verify:setup	check env, Supabase, seed, DICOM storage and OpenAI (-- --env <file>, -- --live-ai)
npm run db:types	regenerate src/lib/supabase/database.types.ts
npm run dicom:generate / dicom:upload	write the phantom DICOM series / upload them to Storage

Layout

src/
  app/                 thin routes: call module use cases only
  lib/                 env (zod), Supabase clients, logger
  modules/<name>/
    domain/            pure types, zod schemas, rules (no frameworks)
    application/       use cases, depend on ports
    infrastructure/    Supabase repositories, LLM adapters, mappers
    ui/                React components
    index.ts           public domain entry point  -> @/modules/<name>
    server.ts          server-only entry point    -> @/modules/<name>/server
supabase/migrations/   schema, RLS, storage (invariants enforced in the DB)
supabase/seed.sql      demo users, templates, synthetic patients and studies
scripts/               phantom DICOM generator and uploader
evals/                 golden cases, scorers and the eval runner (npm run eval)
e2e/                   Playwright specs