Reducing P99 latency in real-time model serving

Proven techniques to shave milliseconds off P99 latency for production model serving — profiling, dynamic batching, compilation, and SLO-driven design