Skip to course content
Free LLMOps course

LLMOps for Reliable AI Applications

Module 01 Activity

Scenario

You have been asked whether an existing LLM feature is reliable enough to expand. Nobody can currently answer, because nobody has defined what reliable would mean.

What you build

A written reliability definition for a real system: an end-to-end target, derived stage targets, and a quality bar someone could run tomorrow.

Steps

  1. Run the same question through the system five times and record the outputs. Judge them on substance and note what differs.
  2. Map every stage from request to response and estimate a success rate for each. Multiply them and compare against what anyone assumed.
  3. Set the end-to-end target first, then derive the per-stage targets from it.
  4. Rank the failure modes by what they cost the user, not by whether your metrics can see them. Say which ones are currently invisible.
  5. Write a quality bar naming the set, the dimension, the threshold and the judge. Check that all four are present.
  6. List the failures you are already accepting, each with a mitigation and a residual risk.

Evidence to hand in

Review checklist