Observability for agentic systems
Tool calls, traces, provenance, eval-as-SLI — what changes when the workload is a model deciding what to do next.
Reliability AI Serving Infrastructure Distributed Systems
Working notes on reliability, GPU infrastructure, network fabrics, and inference observability. Long-form, opinionated, and meant to be useful to engineers operating these systems.
Tool calls, traces, provenance, eval-as-SLI — what changes when the workload is a model deciding what to do next.
Lossless Ethernet is not lossless. It is just a different distribution of where the loss lives.
Why tail latency, queue depth, and NCCL hiccups demand SLOs that ordinary web stacks do not.
Or write to me about any of them. I read every message.