LLM Evaluation and Safety in Production

Curriculum

1. 01Production Is DifferentWhy an AI feature that passed its evaluation can still fail after launch, what changes underneath it, and what four months of a live assistant's metrics and incidents look like.15 min
2. 02Building a Regression SuiteDesign the fixed set of test cases every release must pass (covering categories, difficulty, refusals and safety), and work out how precise its scores are, overall and per category.25 min
3. 03Comparing ReleasesCompare candidate releases with the live one on the same cases, count what each fixed and broke, test whether the difference is real, and find the regressions an overall score hides.25 min
4. 04Release GatesTurn release decisions into automatic rules (minimum scores, no regressions in critical categories, cost and latency budgets) that run on every candidate before it can ship.20 min
5. 05Red-TeamingAttack your own assistant on purpose with a catalogue of techniques, measure the attack success rate by technique, release and guardrail, and turn what gets through into fixes and tests.25 min
6. 06Guardrails and ThresholdsChoose the threshold for an input guardrail by weighing harmful messages missed against harmless customers blocked, and check whether it blocks some groups of customers far more than others.25 min
7. 07Monitoring with Control LimitsSet alert limits from each metric's own normal variation, so unusual days stand out automatically, and measure how much sooner Paystream would have found its incidents.25 min
8. 08Feedback and Sampled GradingLearn why thumbs-up and thumbs-down feedback is a weak quality signal, how a small daily sample graded by people measures accuracy properly, and how pooling days turns a noisy sample into an early warning.25 min
9. 09Incident ResponseMeasure how long incidents took to detect and fix, decide severity and first actions in advance, roll back safely, and write a blameless postmortem that leaves the system better than before.20 min
10. 10Final ProjectPlan your final project, the evaluation and safety programme for a live AI assistant, with a release decision, red-team fixes, a fair guardrail, monitoring that would have caught every incident, and an incident playbook.20 min
11. Final assessmentPass it (60% or more) to earn the LLM Evaluation and Safety in Production course completion badge.