Senior AI Engineer - On-Device CV Architecture Review

We're building a mobile app with computer vision running on-device, and before we commit to the full build we want an experienced engineer to review our approach and tell us where it'll break. We have a trained detection model in PyTorch and a plan for getting it into a real-time mobile pipeline. This first contract is a focused architecture review - we want your judgment on the design before any code. The task: - Review our detection model and the planned on-device pipeline (camera feed → inference → output), and tell us what's sound and what won't hold up in production - Recommend the right on-device path - CoreML vs TFLite vs ONNX Runtime - for our target devices, and why - Flag the real risks up front: where accuracy will drop, where latency will bite, what the framerate ceiling realistically is on a mid-range phone - Point out anything in our architecture that will cost us later - model format, quantization strategy, memory, the tracking layer Requirements: - Real experience taking computer vision models into production mobile apps (iOS/Android) with hands-on model work behind it - Strong opinions on on-device runtimes and the size/speed/accuracy tradeoff for a target device Please share a computer vision system you've taken to mobile - and one architecture decision you'd approach differently today. Please write your estimate.