Point the camera at a dish and get it identified, with macros and a real recipe, in about a second — on-device first, cloud only when needed.

what it does

• Two-stage classification: a 2,023-class TFLite (LiteRT) model handles ~95% of cases on-device; low-confidence cases fall back to Gemini Vision.
• On-device first — inference runs locally in a dedicated isolate (throttled 700 ms/frame), so food data stays on the phone.
• Image-hash caching: SHA-256 of the image (hashed in an isolate) keys a nutrition cache — identical photos return instantly.
• Firebase ML model downloader ships model updates without an app-store review cycle.
• Hand-authored adaptive icon + native splash via a deterministic asset pipeline (no codegen packages).

screenshots

source

https://github.com/RivaldoPardede/Aletheia-Food →