Making AI workloads observable by default on Kubernetes

This talk demonstrates how to inject distributed tracing into AI workloads at the platform layer without requiring application changes. A Kubernetes Operator attaches OpenTelemetry instrumentation to running pods, capturing model inference, latency, and outbound API calls automatically. The resulting traces flow through the standard telemetry pipeline and appear alongside existing services in the observability backend.

The session examines the injection mechanism, the data flow, and the operational implications of automatic instrumentation. The focus is not on the AI framework itself, but on how platform teams can enforce consistent observability standards for emerging workloads through automation instead of policy documents.