Parakh
Parakh (परख, "to assess quality") gives packaged Indian food a single, honest A–E health score. Scan a barcode or photograph the label to get a 0–100 verdict, see why it scored that way, and find healthier like-for-like swaps. You can also snap a home-cooked meal or a thali and track the day against your nutrition targets.
Live: parakh.skdev.one (installable PWA; sign in with Google or continue as a guest)
How it works
barcode ──▶ our DB ──hit──────────────────────────────▶ score + breakdown
│ miss
▼
OpenFoodFacts ──usable nutrition──▶ score ──▶ cache in DB
│ miss / empty
▼
label photo ──▶ vision LLM extracts nutrition + ingredients ──▶ score ──▶ cache in DB
- The LLM extracts, a deterministic function scores. A vision model (via OpenRouter) only reads the nutrition table and ingredient list. Scoring is plain Python: a Nutri-Score backbone (sugar, salt and saturated fat against fibre and protein) mapped to 0–100, minus India-specific penalties for palm oil, maida and additives, plus NOVA-4 ultra-processing markers. The same product always gets the same, explainable score.
- Scored once, cached forever. Every newly resolved product is written back, so the next scan of it is an instant DB hit. Cached rows with no usable nutrition are treated as misses and self-heal on the next scan.
- Healthier swaps by meaning, not keywords. Products are embedded, and alternatives are ranked by cosine similarity, so swaps stay like-for-like (buttermilk → chaas, 70% → 90% dark chocolate). If embeddings are unavailable, it falls back to score order.
- Meals, not just packets. A meal photo is split into individual dishes, each with a typical portion and per-100 g macros. You confirm or adjust the portions, then log it against daily targets for energy, protein, fibre, sugar, saturated fat and salt.
- A seeded catalog. About 1,000 products from Indian e-commerce listings were run through a batch extraction pipeline so Explore and Compare work without scanning first (
docs/catalog-data-pipeline.md).
- Fair quotas. Guests get 3 scans a day and signed-in users get 10. Quota is consumed only on a successful scan, so "needs a photo" or "unreadable label" never burns your allowance.
Stack
| Layer | Technology |
|---|
| Backend | Python, FastAPI, SQLAlchemy 2, SQLite, httpx |
| AI | OpenRouter vision models (label and meal extraction), OpenAI embeddings (swaps) |
| Frontend | React 18, TypeScript, Vite, PWA (vite-plugin-pwa), @zxing barcode scanning |
| Auth | Google Sign-In or guest device tokens, JWT |
| Data | OpenFoodFacts, plus a seeded catalog |
| Hosting | Vercel (frontend), Docker on a GCP VM behind nginx (backend) |
| Quality | pytest (166 tests), Vitest + Testing Library (156 tests), GitHub Actions CI |
Repository layout
backend/
app/
scoring/ nutriscore.py + scorer.py: the deterministic scoring
services/ scan pipeline, meal estimator, auth, rate limiter
clients/ OpenFoodFacts + vision label extractor
repositories/ products, diet log
nutrition/ daily targets
scripts/ catalog build / seed / backfill tooling
tests/
frontend/
src/
screens/ Scan, Result, Compare, Explore, Today, Meal capture, History…
scan/ camera + barcode decoding
api/ session/ diet/ components/
docs/ catalog pipeline runbook, research, design specs
Run locally
# Backend (terminal 1)
cd backend && python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env # set PARAKH_SECRET_KEY; PARAKH_OPENROUTER_API_KEY enables the photo path
uvicorn app.main:app --port 8000
# Frontend (terminal 2)
cd frontend && npm install
echo 'VITE_API_BASE_URL=http://localhost:8000' > .env.local
npm run dev
Open http://localhost:5173.
Tests
cd backend && pytest -q # 166 tests
cd frontend && npm test # 156 tests
CI runs both suites and a production frontend build on every push.
See backend/README.md for the API reference.
