Observability
Tracing, token accounting and evals that make LLM systems debuggable in production — built into my agent work.
Observability is the ability to understand a system from its outputs: traces, metrics and logs. For LLM systems it also means seeing each prompt, tool call, token count and cost, and scoring output quality over time.
How I use it
- Akashi — OpenTelemetry tracing and per-call token usage accounting are built in.
- langgraph-langserve-starter — optional Langfuse tracing on every invoke and stream.
- Conversational AI Agents — Langfuse/OpenTelemetry tracing plus automated Evals in my Independent AI & Full-Stack Engineer (Freelance) work.
- Somnio AI Tools — I added per-step token usage tracking to the CLI's run command.
- mcp-deep-dive — wire-level JSON-RPC frame logging to see exactly what crosses the protocol.
Why it matters
Agents fail in non-obvious ways; traces turn a wrong answer into a findable cause.