Content safety inference at scale

Design the online inference path for a content-safety classifier that must score every user-generated post before it's shown, at 120k QPS with a hard 15ms p99 budget — if scoring is slow it stalls the publish path for real users. The model is a transformer classifier on GPU; naive per-request inference is too slow and too expensive. Walk through how you hit the QPS at that latency, how dynamic batching interacts with the tail-latency budget, and how you degrade gracefully if the model fleet is overloaded rather than blocking every post.