MedFind-Ethiopia/Medfind

MedFind — AI-assisted pharmacy finder for Ethiopia that learns which pharmacy actually has your medicine, using a self-trained Bayesian/Thompson Sampling engine on crowdsourced outcomes (no inventory API needed). Claude API used only for NLU and reply phrasing, never for recommendations. # Medfind Medicine finding for Ethiopia. ## Project structure Single Next.js app (App Router). See the Master Build Blueprint for phase ownership and shared contracts. ``` /app → Next.js app router pages / → landing / consent /chat → Phase 1 chat UI /medications → Phase 2 companion UI /pharmacy → Phase 3 SaaS dashboard (separate auth) /app/api → Next.js API routes /lib /matching → embedding + cosine similarity (Phase 1) /reliability → Thompson Sampling engine (Phase 1) /parsing → instruction extraction (Phase 2) /integrations → Gemini, Addis AI, ElevenLabs, Telebirr /db → Supabase client + query helpers types.ts → shared API + DB types (Section 0.3) /scripts seed.ts → DB seeding /seed-data → curated CSV/JSON sources /supabase /migrations → numbered SQL migrations ``` ## Day-0 setup 1. `npm install` 2. Copy `.env.example` → `.env.local` and fill credentials (optional for UI mocks). 3. `npm run dev` — consent → chat works against local mocks by default (`NEXT_PUBLIC_USE_MOCKS` defaults on). 4. Shared types live in `lib/types.ts` — do not duplicate shapes elsewhere. 5. Work on `feature/*` branches; PR into `dev`, never push directly to `dev` or `main`. ## Frontend (Phase 1) - `/` — consent screen, then chat shell - Consent calls real `POST /api/session` (2t); stores anonymous `userId` + `sessionToken` in localStorage (and a SameSite cookie) - `/api/query` still mockable via `NEXT_PUBLIC_USE_MOCKS` until 3t lands - Brand UI tokens: green ink `#0b6e4f`, warm cream neutrals ## Local environment parity (Section 0.6) | Decision | Choice | Where it is pinned | | --------------- | ---------------------- | ------------------------- …