Nutri-Trust — Smart Pantry

A pantry tracker for Indian households that cuts food waste: scan a barcode or a receipt, see what is about to expire, and get a recipe that uses it up first. Pantries can be shared across a household with row-level security enforcing who sees what.

Live: https://nutri-trusto.vercel.app · installable as a PWA

Screenshot rendered from synthetic pantry data in a test browser, not a real account. There is no demo GIF yet.

What it does

• Scan in — barcode (camera or typed), receipt photo, nutrition label, printed expiry date. Gemini reads the images; Open Food Facts resolves barcodes.
• Know what is expiring — shelf-life estimates with their source shown on the card, risk bands, and a plain "past our estimate, check it yourself" rather than a claim that food has spoiled.
• Cook it — Indian recipes generated from what expires first, validated server-side against the user's diet (vegetarian / eggtarian) before they are shown.
• Share it — households with invite codes, a shared pantry, per-user shopping list and impact dashboard.
• Category hint — a trained model suggests a category for scanned products.

Architecture

flowchart LR
  UI[Next.js PWA] -->|anon key + RLS| PG[(Supabase Postgres)]
  UI -->|bearer token| API[/api routes/]
  API --> GEM[Gemini]
  API --> OFF[Open Food Facts]
  API -.-> ML[FastAPI category model]

The dotted edge is optional. Full diagram, trust boundaries and request flows: docs/ARCHITECTURE.md.

Engineering evidence

Tests & CI	Vitest unit/component/API-route tests, RLS policies run against real Postgres (PGlite), Playwright browser journeys on phone and desktop and a pytest job for the model service. All on every push: workflow
RLS & the 42P17 fix	db/rls.test.ts proves the policies and includes a negative control that does raise 42P17. It also guards the fix for a cross-household write issue the tests found (db/supabase-schema-household-guard.sql)
Responsive & PWA	Before/after measurements in docs/audit/PHASE1_RESULTS.md; a browser test fails the build if a tap target drops below 44px
Load test	docs/perf/LOAD_TEST.md — labelled synthetic, with its limits stated
ML model	ml/README.md — dataset, licence, metrics against the existing keyword rules, calibration, known misses
Audit	docs/audit/PHASE0_AUDIT.md

Stack

• Next.js 16 (App Router, Turbopack), React 19, TypeScript, Tailwind v4
• Supabase — Postgres, Auth, Realtime, row-level security, security-definer functions
• Google Gemini (gemini-2.5-flash) — receipt, label and expiry OCR; recipes; chat
• Open Food Facts — barcode lookup
• scikit-learn + FastAPI — category classifier service
• Vitest, Testing Library, PGlite, Playwright, pytest — tests
• GitHub Actions — CI
• Vercel — hosting

Run it

Locally

npm install
cp .env.local.example .env.local   # fill in Supabase and Gemini values
npm run dev                        # http://localhost:3000

Variable	Used for
NEXT_PUBLIC_SUPABASE_URL, NEXT_PUBLIC_SUPABASE_ANON_KEY	Supabase project (public values)
GEMINI_API_KEY	Scanning, recipes and chat (server only)
ML_SERVICE_URL	(optional) category model service; unset → keyword rules
NEXT_PUBLIC_GOOGLE_SIGNIN	(optional) 1 shows "Continue with Google" once it is set up in Supabase
OFF_USER_ID, OFF_PASSWORD	(optional) contribute barcodes to Open Food Facts

Database

Run in the Supabase SQL editor, in order: db/supabase-schema.sql, db/supabase-schema-additions.sql, db/supabase-schema-hardening.sql, db/supabase-schema-household-guard.sql. All are idempotent. There is no migration ledger; the repo cannot tell you which have been applied to a given project.

Feature migrations follow; each says at its top what must run before it, and the app checks for each column and keeps working without it. The newest, db/supabase-schema-printed-date.sql, keeps the date printed on a pack, so one added after its date reads "10 days ago" rather than "today".

Password reset

"Forgot password?" emails a Supabase recovery link that returns to /reset-password. Two settings in the Supabase dashboard make it work for everyone:

• Authentication → URL Configuration → Redirect URLs: add https://<your-domain>/reset-password. Without it Supabase sends the link to the Site URL instead; the app still forwards it to /reset-password.
• Authentication → SMTP Settings: Supabase's built-in mailer only delivers to members of the project's own team, about two emails an hour. Real users get the email only once a custom SMTP server is set up.

Reminders while the app is closed

A morning push message ("Use these today: Curd, Spinach") for devices that opted in under Notification Preferences. It is off until you set it up, and nothing else changes in the meantime:

1. Run db/supabase-schema-push.sql in the Supabase SQL editor.
2. npx web-push generate-vapid-keys prints a public and a private key.
3. In Vercel, add NEXT_PUBLIC_VAPID_PUBLIC_KEY, VAPID_PRIVATE_KEY, VAPID_SUBJECT (a mailto: address), SUPABASE_SERVICE_ROLE_KEY and CRON_SECRET, then redeploy. The service role key bypasses row-level security: it belongs in Vercel only, never in git or a NEXT_PUBLIC_ name.

vercel.json calls /api/cron/push-reminders once a day at 8:00 am India time. On an iPhone, push works only for the app added to the Home Screen (iOS 16.4+).

Tests

npm run lint && npm run typecheck
npm test                  # unit, component, API-route and RLS tests
npm run test:coverage
npm run test:e2e          # builds, then runs Playwright (phone + desktop + PWA)
pytest ml/service         # model service

Known limitations

• Rate limiting is per server instance, in memory.
• src/app/page.tsx is ~2,900 lines; it is covered by browser tests, not unit tests.
• Statement coverage is about a third of src/ — high where logic lives (lib, API routes), near zero on the big page component.
• The category model is trained on packaged-product names and mislabels fresh produce; it is not wired to anything that spends money.

Layout

src/app/         routes, layout, API routes (/api/health, /api/categorize, …)
src/components/  UI components
src/lib/         matchers, shelf-life data, auth, rate limit, upload validation
db/              SQL migrations; db/proposed/ holds reviewed-but-unapplied ones
e2e/             Playwright specs and the fake backend they run against
ml/              data fetch, training, evaluation, service, committed model
perf/            Locust load test and stub
docs/            architecture, ADRs, audits, perf report

FEATURE_IDEAS.md tracks planned and shipped features; CHANGELOG.md records every change to shelf-life data with the reason.