LLM Evaluation and Safety in Production

Programme overview

Launch is where the work starts. In this course you run quality and safety for Paystream's help assistant after four months in production and three incidents. You'll build and size a regression suite, compare two candidate releases case by case and find the fraud regression an overall score hides, and turn release decisions into automatic gates. You'll measure 240 red-team attacks by technique, release and guardrail, choose a guardrail threshold from stated costs and discover that it blocks harmless Pidgin messages far more often than English ones. Then you'll set alerts from each metric's own variation, see why thumbs-down feedback missed an accuracy drop that pooled graded samples catch, measure time to detect, write a blameless postmortem, and backtest a monitoring plan against every incident. Every number comes from running the code.

Curriculum

1. 01Production Is Different15 min

   • Production 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.
2. 02Building a Regression Suite25 min

   • Building 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.
3. 03Comparing Releases25 min

   • Comparing 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.
4. 04Release Gates20 min

   • Release 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.
5. 05Red-Teaming25 min

   • Red-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.
6. 06Guardrails and Thresholds25 min

   • Guardrails 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.
7. 07Monitoring with Control Limits25 min

   • Monitoring 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.
8. 08Feedback and Sampled Grading25 min

   • Feedback 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.
9. 09Incident Response20 min

   • Incident 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.
10. 10Final Project20 min

    • Final project: Paystream's AI quality and safety planPlan 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.

Assessment and certificate

You are assessed with a final assessment (pass mark 60%) and a final project. Your certificate is awarded when you meet those requirements. Paying for the programme does not award it on its own.