Wiring Android's ML Kit Barcode Scanner to a On-Device LLM for Real-Time Product Intelligence

Chaining ML Kit's barcode/QR detection pipeline directly into MediaPipe LLM Inference API — covering frame throttling between vision and language stages, shared GPU memory to avoid CPU round-trips, structured output constraints for product JSON, and the end-to-end P95 latency ceiling (camera → barcode → prompt → first token) that separates a snappy UX from a frustrating one