AI Product Engineering
Block 3: Agents, Evaluation & Deployment
Traditional testing: same input → same exact output.
AI testing: same input → probabilistic output — "correct" is a matter of degree.
Different framework required — not a different level of rigour.
Is the answer factually accurate? Supported by context?
Does it actually answer the question asked?
Same question, different phrasing → same answer?
The model is not lying — it is predicting the most likely token sequence. The result can be fluent, confident, and completely wrong.
Choose the best diagnosis.
Choose the best diagnosis.
Choose the best diagnosis.
| Category | Tests |
|---|---|
| Out-of-scope | Does it refuse or hallucinate? |
| Ambiguous | Does it clarify or guess? |
| Contradictory | Does it maintain consistency? |
| Edge case | Empty, very long, special characters |
| Injection | "Ignore all previous instructions…" |
True or false — cold-call a team.
True or false — cold-call a team.
True or false — cold-call a team.
test-suite.example.json → expand to 10 cases in-sessionnpm run eval → score 1–5 correctness & relevanceEvaluation-Report.mdShowcase gate remains ≥15 cases — finish remaining in Stretch / Session 15 prep.
Next Session: Security, Privacy & Responsible AI