Design an ML System to Detect Dangerous Driving

Waymo wants a production-grade machine-learning system that can continuously flag dangerous driving situations in real time from the perception and planning stack of its autonomous vehicles. Your design must (1) define what "dangerous" means and how ground-truth labels are generated at fleet scale without human reviewers in the loop, (2) choose features, model architecture, and training infrastructure so the detector runs on every vehicle at 10 Hz with <50 ms added latency, (3) guarantee that the model maintains ≥99 % recall on the most severe 0.1 % of frames while keeping false-positive rate below 5 %, (4) detect and react to distribution drift within 24 h even when fresh labels are not yet available, and (5) provide a continuous-learning pipeline that safely retrains and validates new models every week on 10 PB of driving data. You should sketch the end-to-end data flow from on-car perception outputs through offline label generation, training, evaluation, canary rollout, and automatic rollback. Be prepared to justify every design choice with safety, scale, latency, and fleet-update constraints.