SoloHard40mPro Scenario
The Assistant That Cited the Wrong Policy
A support assistant gives an EU Pro customer a 30-day cancellation window and cites a real policy. The current policy says 14 days. Investigate the RAG trace, contain unsafe policy answers, repair metadata-aware evidence selection and abstention in GitHub, replay a fixed canary audit, and write the post-mortem.
Skills you'll prove
ai-engineeringragretrieval-evaluationmetadata-filteringgrounded-generationabstentionincident-response