Introduction: Inference serving has evolved from a research afterthought into the single most expensive and failure-prone component of production AI systems. As Abi Aryan highlighted in her LLMday keynote announcement, GPU economics, latency tails, and silent failure modes are now first-order reliability problems—not edge cases—demanding rigorous Site Reliability Engineering (SRE) principles tailored for LLM and agentic workloads. Learning Objectives: Diagnose and mitigate silent inference failures (model drift, hardware degradation, memory leaks) using observability stacks and chaos engineering.

Conference on Large Language Models, AI and Machine Learning. Probably the most awesome LLM conference in Harness, San Francisco, United States!