Skip to content

PHASES.md

Latest commit

 

History

History
132 lines (91 loc) · 4.78 KB

PHASES.md

File metadata and controls

132 lines (91 loc) · 4.78 KB
Raw

NutriLens — Development Phases

Full build log from first commit to production. Updated as phases complete.

✅ Phase 1 — Chrome Extension Foundation

March 2, 2026

Chrome MV3 extension scaffold with React popup, content script, and service worker. Establish DOM extraction and build pipeline.

  • Chrome MV3 extension scaffold — Vite + React
  • amazon.js content script — DOM extraction of name, brand, price, quantity
  • background.js service worker — session storage, message routing
  • Popup UI skeleton
  • Build pipeline (npm run build → chrome://extensions)

✅ Phase 2 — Backend Infrastructure

March 3, 2026

Stand up FastAPI with PostgreSQL, Redis, and Celery. Define full database schema.

  • FastAPI app with async SQLAlchemy
  • PostgreSQL schema — Products, NutritionFacts, ExtractedClaims, Contradictions, ProductScores, AnalysisJobs
  • Redis for cache (30-day TTL) + Celery broker
  • Docker Compose wiring all services
  • CORS configured for Chrome extension origin

✅ Phase 3 — Extension ↔ Backend Wiring

March 3, 2026

Connect the extension to the backend end-to-end.

  • Extension POSTs product payload on every page load
  • Celery worker skeleton for async analysis
  • Job polling — submit → job_id → poll every 2s → SCORES_READY broadcast
  • MV3 service worker persistence fixes
  • Price parser fixes for Indian number format

✅ Phase 4 — OCR Pipeline

March 3–4, 2026

Separate OCR microservice with PaddleOCR. Two-pass pipeline for accurate per-100g nutrition extraction from label images.

  • Separate OCR microservice in Docker
  • Two-pass pipeline — Tesseract PSM11 keyword scoring → PaddleOCR spatial extraction
  • Coordinate-based spatial parser using bounding box x/y positions (outperforms line-by-line)
  • Per-100g column detection, serving range sanity validation, x-cap for amino acid tables
  • Snippet capture UI — drag-select overlay on Amazon's HD image viewer
  • Base64 image transfer, threading lock for PaddleOCR concurrency
  • Benchmark: 11/11 correct nutrients on reference label image

✅ Phase 5 — Data Quality & FSSAI

March 4, 2026

FSSAI licence extraction, NutriScore, and column detection improvements.

  • FSSAI extraction from seller About page — DOM fetch, regex, 14-digit state code validation
  • Semantic parser column detection fixes — relative x_cap threshold (0.85 × image width)
  • NutriScore A–E computed per-100g per FSSAI thresholds
  • CompareTable with Protein/₹100 as primary comparison dimension
  • Per-100g adopted as canonical unit throughout

✅ Phase 6 — Accountability Layer ← current

March 7, 2026

Full claims pipeline — extract, store, classify, verify, score. Admin dashboard.

  • Structured claims extraction — each bullet with DOM selector and element index
  • Claims saved to PostgreSQL on every page load (deduplicated)
  • Clickable claims — tap to scroll + yellow highlight on product page
  • Certification links — Labdoor, Informed Choice, Informed Sport, Trustified, NSF
  • Numeric claim checker — cross-validates claimed values vs label with 15% tolerance
  • Accountability score — value (35%) + quality (30%) + integrity (35%)
  • ClaimVerification table — verdict per claim: verified / contradicted / unverifiable
  • Normalizer bug fixed — per-100g OCR values no longer double-multiplied
  • Admin transparency dashboard — live Postgres + Redis browser, truncate/delete/flush

🔲 Phase 7 — LLM Claim Intelligence

Upgrade claim classification from regex rules to LLM reasoning.

  • Groq free tier (llama3-8b) or Ollama (local) in the Celery worker
  • Deep claim classification — efficacy, comparative, superlative, ingredient sourcing
  • Integrity score upgraded to LLM-verified
  • Study citation extraction for efficacy claims ("clinically proven", "50% faster absorption")
  • Contradiction detection — flag when label data contradicts marketing copy

🔲 Phase 8 — Multi-Platform

Extend content scripts to BigBasket and Flipkart.

  • BigBasket content script — category page + product detail
  • Flipkart content script — nutrition table format differs from Amazon
  • Platform-agnostic extractor interface in background.js

🔲 Phase 9 — Product History & Discovery

Persistent view of everything the user has scanned.

  • GET /api/v1/products/history — paginated from PostgreSQL
  • Third tab in popup — Previously Scanned with search and filter
  • Category leaderboard — best accountability score per category
  • Price drop alerts for tracked products

🔲 Phase 10 — Production

Harden, deploy, publish to Chrome Web Store.

  • Auth on write endpoints (API key or JWT)
  • Rate limiting on /products/submit
  • Production deploy — Railway / Render / VPS
  • /admin in read-only mode for public access (ADMIN_READ_ONLY=true)
  • Chrome Web Store submission