Julienne — shipped iOS cooking app

An iOS app on the App Store. Point the camera at a grocery receipt and it becomes your pantry; then an assistant named Juli suggests what to actually cook with what you have, tracks what you made, and learns your preferences.

The hard part is the receipt. Grocery receipts are close to worst-case OCR: thermal print, abbreviations, inconsistent casing, and store-specific codes — GV WHP CRM 8Z has to become “whipping cream” before anything downstream works. That’s an OCR pass feeding a normalization layer (OCRService → ReceiptNormalizer → IngredientNameNormalizer), with a streaming review screen so the user corrects the model rather than waiting on it. Same problem shape as production document extraction, in a consumer app where a wrong line item is immediately obvious to the user.

The assistant is grounded rather than freeform: it disambiguates when a request is underspecified, confirms preferences instead of assuming them, and shows its lineage on a suggestion so a recommendation can be traced back to what’s in the pantry.

Shipping it was the other half. StoreKit 2 subscriptions with a paywall and App Store Server API validation, Sign in with Apple, a Supabase backend with cross-device sync, a written privacy policy, event instrumentation as a build-time habit rather than an afterthought, and GitHub Actions pipelines for the iOS build, the backend, and the marketing site.

Private repo — the portfolio entry is a write-up. The app itself is public on the App Store.