A terroir-first, privacy-respecting wine journal and recommender.

• Web: Next.js 16, TypeScript, Tailwind CSS, shadcn/ui
• Backend: Supabase (PostgreSQL 15, Auth, Edge Functions, Storage, pgvector)
• iOS: SwiftUI (iOS 17+), MapKit
• Android: Kotlin, Jetpack Compose, Material 3
• Secrets: Doppler
• CI/CD: GitHub Actions, Vercel
• AI: OpenAI GPT-4o (label extraction), Jina CLIP v1 (visual embeddings), Supabase gte-small (text embeddings)

• Node.js 22+
• pnpm 11+
• Docker (for local Supabase)
• Supabase CLI: brew install supabase/tap/supabase
• Doppler CLI: brew install dopplerhq/cli/doppler

# Install dependencies
pnpm install

# Start local Supabase (Postgres, Auth, Storage, Edge Runtime)
supabase start

# Reset database, apply migrations, seed data
supabase db reset

# Sync API keys from Doppler for local edge functions
./scripts/sync-env-local.sh

# Start the dev server
pnpm dev

Open http://localhost:3000.

All secrets are managed in Doppler under the vinho project. The pnpm dev command injects them automatically via doppler run.

Edge functions run inside the Supabase Docker runtime and can't use doppler run directly. Instead, supabase/.env.local bridges Doppler secrets to the edge runtime:

# Generate supabase/.env.local from Doppler
./scripts/sync-env-local.sh

# Serve edge functions locally
supabase functions serve --env-file supabase/.env.local

This file is gitignored. Run sync-env-local.sh whenever you rotate keys in Doppler.

Seeded automatically by supabase db reset:

This user has 20 tastings, 3 scans, and 1 pending queue item seeded for testing.

pnpm dev                    # Start Next.js dev server (injects Doppler secrets)
pnpm lint                   # Lint the web app
pnpm --filter vinho-web run typecheck  # TypeScript check

pnpm --filter vinho-web run test              # Unit tests (no Supabase needed)
pnpm --filter vinho-web run test:integration  # Integration tests (requires local Supabase)
pnpm --filter vinho-web run test:ci           # All tests for CI

pnpm simulate               # Upload 10 wine label images and queue them
pnpm simulate:process        # Same, plus invoke the process-wine-queue edge function

The simulation uploads real wine label fixtures from supabase/fixtures/wine-labels/ to local storage, creates scan records, and queues them for processing.

supabase start              # Start local Supabase
supabase db reset           # Wipe and rebuild from migration + seeds
supabase migration new <name>  # Create a new migration
supabase db push            # Push migrations to production (or via CI)
supabase gen types typescript --local > apps/vinho-web/lib/database.types.ts  # Regenerate types

supabase functions serve --env-file supabase/.env.local   # Serve all functions locally

vinho/
  apps/
    vinho-web/              # Next.js 16 web app
    vinho-ios/              # SwiftUI iOS app
    vinho-android/          # Jetpack Compose Android app
  supabase/
    migrations/             # Production schema (single pulled migration + incremental)
    migrations-archive/     # Old local-only migrations (preserved for reference)
    functions/              # Edge Functions (process-wine-queue, generate-embeddings, etc.)
    fixtures/wine-labels/   # 10 real wine label images for simulation
    seeds/                  # Seed SQL source files
    seed.sql                # Active seed (test user, 20 wines, tastings, scans)
    config.toml             # Local Supabase config (ports, buckets, auth)
    .env.local              # Edge function secrets (gitignored, generated from Doppler)
    shared/                 # Shared utilities for edge functions
  scripts/
    sync-env-local.sh       # Sync Doppler secrets to supabase/.env.local
    simulate-wine-upload.ts # Wine upload simulation (lives in apps/vinho-web/scripts/)
  .github/workflows/
    ci.yml                  # PR checks: lint, typecheck, unit + integration tests
    deploy.yml              # On merge to main: push migrations, deploy edge functions

GitHub Actions runs lint, typecheck, and both unit and integration tests (with local Supabase) on every PR to main.

On merge to main:

1. Supabase migrations are pushed via supabase db push
2. Edge functions are deployed via supabase functions deploy
3. Vercel auto-deploys the web app

When a user scans a wine label, Vinho runs a multi-step pipeline that identifies the wine, creates database records, and builds a searchable index for future scans.

1. Image capture: Client compresses to max 2000px JPEG at 0.8 quality
2. Upload and queue: Image goes to Supabase Storage scans bucket, scan record created, queue item inserted as pending
3. Visual embedding match (fastest): Jina CLIP generates 768-dim embedding, searches wine-labels vector bucket. Threshold: 92% similarity
4. Text vector match (free): Supabase gte-small generates 384-dim embedding from OCR text, searches wine_embeddings via pgvector. Threshold: 90% similarity
5. OpenAI Vision extraction: GPT-4o-mini extracts structured data. Below 60% confidence, escalates to GPT-4o
6. Record creation: Region, producer, wine, vintage, and grape varietals created with upsert logic
7. Embedding storage: Visual embedding stored for future matches, text embedding queued
8. Enrichment: Separate queue fills in tasting notes, food pairings, aging potential via GPT-4o-mini

Private. All rights reserved.