Skip to content
Mariano Rodrigo

Mariano Rodrigo

AI Solutions Engineer building production systems with artificial intelligence, automation, and full-stack architecture. This is my public engineering lab: architecture decisions, implementation reports, experiments, and lessons from real systems.

Observability for AI workflows

· observability, ai-systems, operations

Traditional observability watches CPU, latency, and errors. AI workflows add a harder question: was the output any good — and that answer is not in your metrics dashboard by default.

What to trace per request

  • Inputs and the exact prompt sent
  • Retrieved context and its sources
  • Model, version, and parameters
  • Tokens, latency, and cost
  • The decision taken and whether a human reviewed it

Why audit trails matter

When a stakeholder asks “why did the system say this?”, a stored trace turns a panicked investigation into a lookup. For regulated or high-trust domains, that record is the difference between a system you can defend and one you cannot ship.

Operating, not just monitoring

Aggregate the per-request data and you get the signals that actually drive improvement: retrieval recall over time, hallucination rate by document type, cost per successful task, and the share of outputs that needed human correction.

Observability is what converts an AI prototype into something you can run on purpose.