noor issa

Skincare labels are long lists of names most people can't decode. skinscript is a full-stack beauty-tech site that explains an ingredient list in plain language and flags common conflicts across a whole routine, with a cited source behind every flag. Paste a list, snap a photo of the label, or search for a product by name or barcode.

A Next.js and TypeScript app. Tesseract.js reads label photos in the browser, so they never leave the device. A parser normalizes ingredient names, including OCR misreads, and a rules engine checks the routine against a small knowledge base where every entry cites a source. Product search uses the Open Beauty Facts open database. Optional Supabase accounts add saved checks, "my routine", an ingredients-to-avoid list and a moderated community chat, all protected by row-level security in Postgres. An optional AI summary is grounded only in the curated data.

• Every flag links back to a cited source
• Six languages, including right-to-left Arabic
• Photos are processed on-device and never uploaded
• 89 tests passing, run on every push

I'm always hunting for cafés with the best matcha and coffee, so I built a place to log every visit, rate the matcha and the coffee separately, see it all on a map, and get an honest answer to "what should I try next?"

A Next.js 16 (App Router) app in TypeScript with Prisma on PostgreSQL, and SQLite for zero-setup local dev. The recommender is plain, tested TypeScript: tasting notes become a TF-IDF taste profile weighted by rating, untried cafés are ranked by cosine similarity with a small neighborhood nudge, and the terms that moved each score become the reasons shown on screen. An optional LLM summary sits on top and the page works fully without it.

• Every recommendation explains itself, instead of being a black box
• Starts with 10 real cafés across Boston, Cambridge, Somerville and Brookline, each linked to its Google Maps reviews
• Leaflet map on OpenStreetMap, so no paid API keys
• Live demo runs in the browser on GitHub Pages, with anything you add saved only on your device
• CI runs lint, typecheck and tests, then builds against a real PostgreSQL service

Long walks are my favorite way to spend a free day, so I built the kind of pipeline I work with professionally around walking and activity data: raw exports in, clean and trustworthy insights out.

Raw files land and move through bronze, silver and gold layers in DuckDB. Bronze enforces a column contract and deduplicates files by SHA-256. Silver casts types, quarantines bad rows with the reason instead of dropping them, and upserts incrementally behind watermarks, with data and watermark committed in one transaction. Gold is rebuilt and quality-checked atomically, so a failed check rolls back and the dashboard keeps the last good tables. Importers handle Apple Health and Google Fit exports.

• Nothing disappears silently: every rejected row lands in quarantine with its reason
• A failing quality check never reaches the dashboard
• One command builds every layer and serves a Streamlit dashboard that sums up the data in plain English, on synthetic data by default