← CurriculumWeek 08 / 10 · Quality
08
Evals: Knowing Your AI Actually Works
Stop shipping vibes. Start shipping measured outputs. This is the single biggest discipline that separates client work from playing around.
Topics covered
- 01Why evals matter — especially for client-facing AI work
- 02Types of evals: deterministic, LLM-as-judge, human review
- 03Writing test cases and building a golden dataset
- 04Iterating prompts and systems using eval feedback (the inner loop)
- 05Eval tooling: Braintrust, LangSmith, or a simple Python harness
- 06Wiring evals into client deliverables and acceptance criteria
- 07Eval anti-patterns: cherry-picking, judge-model bias, vanity metrics
Build challenge
Deliverable
Eval suite (5–10 cases) for the Week 7 agent. Iterate the agent based on results. Document the deltas.