## Question Design an end-to-end system to detect and moderate harmful videos on a large platform. ## Requirements - Detect multiple policy categories (violence, self-harm, hate, sexual content, etc.). - Operate at upload time and for already-published content. - Combine automated decisions with human review. - Minimize both false negatives (missed harmful content) and false positives (wrongful removals). - Provide appeals and auditability. ## Deliverables Architecture, model strategy (multi-modal), labeling/review workflow, metrics, and how you handle adversarial evasion.