On-device model deployment and telemetry

Design the system that ships and runs an on-device computer-vision model on 50M smartphones (e.g. a real-time camera segmentation feature). Inference runs locally for privacy and latency, but the model must be updatable over the air, must work on a wide range of device capabilities, and you need to know in aggregate whether a new model version is actually better — all without uploading users' images. Design the on-device inference + update + telemetry system.