AI Engineering Cumulative Review

Integrated scenario

An organization wants an assistant that answers policy questions and may draft—not approve—support actions. Sources change daily, contain access restrictions, and sometimes conflict.

Produce:

1. decision contract and non-AI baseline;
2. dataset/source provenance and deletion design;
3. model or retrieval architecture with rejected alternatives;
4. frozen retrieval, answer-quality, safety, subgroup, latency, and cost evaluation;
5. threat model covering prompt injection, exfiltration, poisoning, and tool misuse;
6. system version manifest and progressive-release plan;
7. SLOs, fallback, rollback, escalation, and incident drill;
8. AI RMF risk register and residual-risk decision.

Oral synthesis

Defend one claim from raw evidence, diagnose one injected failure, reject one inappropriate AI feature, and explain how a correction propagates through evaluation and production.