Building audit-grade AI agents in the open.
Flagship:
Pre-Approvals Reviewer — an agentic verification pipeline for purchase pre-approvals in a Medicaid-audited NY disability program (github.com/ebt55/preapproval-verification-challenge).
• Five-stage pipeline: vision PDF extraction → category-routed YAML checklists → deterministic fee-cap/eligibility checks → browser research agent (Anthropic SDK + Playwright) → HTML/JSON reports with screenshot evidence.
• Integrity gates make fabricated evidence structurally impossible: "Found" requires a real on-disk screenshot, quotes must appear verbatim on a visited page, and timestamps/URLs/SHA-256 hashes are written by code the model never touches.
• Validated in layers: 18 deterministic tests, a rerunnable hash-manifest audit, human ground-truthing, and planted-trap cases the agent caught.
