A. Document-to-insight (full-stack, front-end-strong): ingest a financial document (statement or factsheet), extract it to a validated schema, and render a polished, responsive, accessible React summary - with reliability guards (timeouts, retries, graceful degradation), an explicit “never auto-act” boundary, and a visible “verify before you act” cue.
B. Multi-model trust (orchestration): route one prompt to two or more models, show where they agree and disagree, attach a source/citation to each claim, and validate the structured output before display. Include an eval measuring disagreement detection or citation correctness.
C. CI & release (platform): stand up a CI pipeline + minimal IaC for an agent service - tests, an eval/quality gate that catches a regression in agent output, and a one-command rollback. Show the pipeline going red on a bad change.
D. Hardening & observability (security/privacy): harden a deliberately naive deployment - least-privilege access, secrets management, no-real-data fixtures, request-ID tracing, plus one alert on a meaningful failure. Treat the data as sensitive financial PII.
What to submit
- A Git repo with a 10-minute README.
- A test or eval harness (for the platform brief, show the pipeline failing on a bad change).
- A one-paragraph AI-use disclosure: what Claude or Codex drafted, and what you hand-verified.
- For front-end briefs, a deployed link, or screenshots of the key states.