Privacy-first cosmetic ingredient analyzer. Users photograph an INCI label and get an AI-generated risk score, comedogenic rating and irritation potential per ingredient — plus cross-product compatibility analysis and a drag-and-drop weekly skincare routine. User data (history, calendar, profile) lives only on the device; the server keeps just anonymous cache + rate-limit metadata.
Cosmetic ingredient lists (INCI) are opaque: most consumers cannot interpret the names, nor detect conflicts between products used simultaneously. Skinwise closes that gap with a multimodal AI pipeline that stays cheap, fast and privacy-respecting.
How it works
Users upload 1–4 photos of a product label — multi-photo support handles cylindrical packaging where the full list wraps around. The server normalizes images via sharp to ≤1024px JPEG, computes a SHA-256 hash, looks it up in Upstash Redis (24-hour TTL), and on cache miss calls Gemini 2.5 Flash-Lite (multimodal vision) with a structured Turkish-language prompt. The API returns a typed ProductAnalysis object with per-ingredient breakdowns and a composite risk score, then writes the result back to Redis so a re-upload of the same image hits the cache for free.
Routes
/icerik-analizi — single-product label analysis
/cakisma-analizi — accepts 2–3 products, surfaces RoutineConflict records (ingredient pairs that should not be used together)
/rutin — drag-and-drop weekly skincare routine; old /takvim URLs permanent-redirect here. The backend endpoint stays /api/routine.
Privacy & data architecture
User-facing data (analysis history, calendar entries, profile) lives only on the device — localStorage with a 30-entry rolling cap, no account, no upload. The server only persists two anonymous artifacts in Upstash Redis: an IP hash for rate-limit counters, and an image-hash → ProductAnalysis cache (24-hour TTL, no user identity attached).
A storage.ts migration layer auto-splits legacy multi-day routine entries into per-day records, so each day has its own independent completion checkmark. Analyses can be exported as PNG report cards via a Canvas-based renderer (lib/export-report.ts).
Smart cache — same image re-uploaded inside 24h costs $0 (Redis hit, no Gemini call)
Theme-aware brand — light/dark logo + favicon driven by prefers-color-scheme
Drag-and-drop routine planner — move routines between days, Alt+drag to copy
Per-day completion checkmark — independent done-state per day
Multi-day storage migration — legacy multi-day entries auto-split into per-day records
Cosmetic guard, 3-layer defense — prompt directive + server-side override + frontend safety net (the model's freeform image commentary is never shown to the user)
Cause-aware cooldown banner — distinct Turkish messages for Gemini rate-limit vs. Gemini overload vs. app rate-limit
