Global social platform reaches 99.6% moderation accuracy on Amazon Bedrock

Content moderation is a classification problem with asymmetric error costs: a false negative puts unsafe content in front of users; a false positive merely queues a human review. That asymmetry dictates the whole ML program — recall is the decisive metric, and any model change must be judged against a stable labeled benchmark, not anecdotes. The platform also faced genuine model uncertainty: multimodal vision-language models were improving monthly, and the incumbent classical-CV engine’s accuracy and unit economics both needed to be beaten with evidence before production could move.

VeUP built the evaluation machinery first. A 1,501-image labeled benchmark became the standing ground truth, and Phase 1 ran a controlled head-to-head across candidate engines — multimodal foundation models on Amazon Bedrock (Qwen 3 VL, Llama 4 Maverick, Amazon Nova Premier, Claude Opus) against the classical computer-vision baseline — with every candidate scored on recall, precision behavior, latency, and cost per 1,000 evaluations. Qwen 3 VL won on the decisive metric and unit economics, and Phase 2 promoted it to the production engine with the baseline retained as a fallback. Video classification composes per-frame inference into a content-level verdict through a K-of-N aggregation rule, with AWS Step Functions pacing per-frame inference and AWS Lambda executing it. In production, Amazon CloudWatch tracks the ML program’s health as measured quantities — recall against the benchmark, P50/P95/P99 latency, and cost per evaluation — so the bake-off can be re-run and the engine re-selected as the model landscape moves.