MedLens

Problem Definition

Most people in India know only the brand name of their medicine. They don't know the active ingredients, don't know whether a cheaper brand has the exact same composition, and have no easy way to start from a photo of what's in their hand. Existing pharmacy platforms show alternatives, but none let you point a camera at the strip and get a straight answer.

Implementation Role

• Built the full app end-to-end as a solo developer — design, backend, frontend, data pipeline, and deployment.
• Wrote a 3-engine OCR pipeline: Gemini Vision reads the photo server-side, while PaddleOCR and Tesseract run in the browser so the server never sees the image.
• Added a recovery step that matches OCR output against the real product catalogue to pair each dose with the right ingredient when the layout gets scrambled.
• Built a matching engine that turns normalised ingredients, strengths, and dosage form into a deterministic composition key, then finds equivalents with a single indexed lookup across a 250K-row catalogue.
• Used PostgreSQL trigram search (pg_trgm) with a 0.85 similarity threshold for fuzzy ingredient matching — strict enough that ‘folic acid’ is not mistaken for ‘valproic acid’.
• Added non-dismissible safety warnings for Narrow Therapeutic Index drugs and Schedule H/H1/X medicines.

System Architecture

• Frontend: React 19 SPA served as static assets from Cloudflare’s edge.
• API: Hono running on Cloudflare Workers.
• Database: Neon PostgreSQL with Prisma ORM, connected through Cloudflare Hyperdrive for pooled connections.
• Rate limiting: two tiers — 60 requests/min for all APIs and 10 requests/min for scans, which trigger the expensive Gemini calls.
• Vision pipeline: Gemini → PaddleOCR → Tesseract fallback chain.
• Privacy: anonymous scan event logging with no personal data, in line with the DPDP Act 2023.
• Data: Kaggle A–Z medicines dataset, the 1mg catalogue, and the official Jan Aushadhi (PMBI) price list.

Challenges & Solutions

• Real-world photos — angled, blurry strips with foil glare — made OCR unreliable.→ Gemini Vision as the primary reader, since it handles these photos well.
• Gemini costs money for every image.→ A fallback chain where PaddleOCR runs free in the user’s browser.
• PaddleOCR often read ingredient names correctly but mixed up which dose belonged to which ingredient.→ A catalogue recovery step that checks which combinations are actually sold in India.
• Fuzzy search needed a careful threshold — too low risks confusing dangerous drug names, too high rejects genuine typos.→ A 0.85 trigram similarity threshold.

Deployed Impact