Built around one rule: when the analyzer fails, the user sees a retryable error, never a plausible fake plate. Nutrition apps that guess silently are worse than no app.

MealLens is a Capacitor-wrapped React app for iOS. Photograph a meal and it returns per-food calories, protein, carbs, fat, and a confidence score to confirm or edit before logging. There is no backend and no account; meal history lives on the device and only the photo leaves it.

Analysis runs through a preference-ordered chain: Apple's on-device FoundationModels first (wired in, not yet serving in the TestFlight build), then a self-hosted Qwen2.5-VL vision model on an RTX 3090. A backend failure surfaces as an error rather than a substituted result, a rule adopted after an early build turned a grapefruit into spaghetti. A separate evaluation harness scores the on-device model against 15 real, weighed meals before it is trusted as the primary path.

• → Land the on-device model on a real iOS 27 device
• → App Store submission decision

• ! The GPU fallback depends on a home workstation being reachable