03 / 05 · SIA Learn and Hack 2026 · Semi-finalist
DocReview
The model flags the gaps. A person makes the call.
Python · FastAPI · Anthropic SDK · pypdf · SQLite
The problem
Teams swap vendor security documents and contracts by email and read them line by line against policy. Each team uses AI its own way, so reviewers can’t see or control how a document was checked.
My role
Team hackathon entry. I built the full working app and the pitch deck, and pitched it with the team.
How it works
- ConfigReviewer sets up portalSets system prompt and baseline policy
- InputSubmitter uploads documentTXT, PDF or DOCX converted to text
- PromptPrompt is assembledBaseline, document and JSON rule merged
- ModelLevel 1 AI reviewClause gaps, each with a risk rating
- Code gateJSON parsed and loggedFailures logged, never shown as results
- HumanLevel 2 decisionApprove, or request a new version
Model stepHuman or code gatePlain step
What I built
- 01Built the end-to-end app and the pitch deck for a team entry that reached the semi-finals
- 02Built a FastAPI REST API with SQLite and a single-file frontend, with data that persists between runs
- 03Added versioned resubmission, a version diff, Word export and an audit trail of every action
Trust boundaries · where the AI is held in check
- AOnly the reviewing team controls the system prompt and baseline, so submitters can’t change what the AI checks against.
- BThe AI never closes a submission; a person’s Level 2 decision does.
Outcome
Semi-finalist at SIA Learn and Hack Hackathon 2026, and well received by the product owner in user acceptance testing.
Screens
