Paper Trail Agent
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote
Paper Trail Agent checks loan and bank statement PDFs for contradictions, so someone questioning a lender's numbers can bring exact evidence rather than a guess. It extracts principal, interest, dates and balances, recomputes totals with Python Decimal, and links findings to highlighted original source lines. Matching account, currency and period checks prevent invalid comparisons. A separate Extractor, Verifier and Critic validate the evidence; uncertain scans request review. Saved human decisions, exports, bounded tool traces and visible evals make the result auditable. Duplicate PDFs cannot be compared against themselves. Try the protected synthetic demo: https://paper-trail-agent.vercel.app/ . It shows a $25 cross-statement interest discrepancy, a stated total $50 too high, a clean control, and an image-only review case. Public guests use deterministic checks with zero paid model calls; separate published live Muse Spark tests verify actual Meta tool use. The public repo includes fictional known-answer data, 407 passing core tests and CI. Real-world layout accuracy is not claimed; hosted OCR worker, email/portal ingestion and continuous monitoring remain pending. This is an evidence and review tool, with no automatic payments or dispute sending.