Countersign

Inspiration

This summer I worked at Blaze, a YC payments startup. Bills landed from everywhere, many for thousands of dollars, and startups don't have an accounts-payable team. They have whoever's free. That's exactly when a fake invoice gets paid. Business email compromise cost US businesses about $3B in 2025 (FBI IC3).

What it does

Countersign checks every bill before it's paid: from your Gmail inbox automatically, a PDF or phone-photo upload, a CLI, or an API. It runs 14 checks in three groups (bill, email, document), scores the risk, and stamps STOP, HOLD or CLEAR with evidence on every flag. It never pays. A human signs, and every "false alarm" or rejection re-tunes the weights for that company. Our mascot Sig has tiny arms: he can stamp a bill, but he can't press pay.

How we built it

• Reading: pdf.js text layer, Tesseract OCR for photos, a field parser. No AI model in the verdict.
• Evidence (6 sources, in parallel): vendor payment history, Glasser (Hunter company records + Serper scam-report search), RDAP domain age, Gmail DMARC/SPF/DKIM headers, PDF metadata.
• Checks: lookalike-domain detection (homoglyph normalization + edit distance), bank-account change, robust z-score amount outliers (Iglewicz–Hoaglin), weighted pressure-tactic phrases, near-duplicate invoice numbers, ABA/IBAN checksums, line items vs total, PDF fingerprint vs the vendor's past bills.
• Scoring: weighted points + combination bonuses, a confidence measure, and STOP/HOLD/CLEAR bands. Email and document checks can only make a stamp stricter.
• Node server + static web app, CLI that exits with code 2 on any STOP so a payment run refuses to go out.

Challenges

Real invoices are messy: card-paid bills with no bank account, address formats that differ, billing platforms sending on behalf of merchants. Each needed a rule that reports "unknown" instead of guessing.

What we learned

The AI shouldn't make the call. The value is in making a person's judgment fast and well-informed.

What's next

Accounting-tool plugins (pay button respects the stamp), vendor call-back verification, and per-company learning across a team.

Credits: architecture ideas from Reva (MIT) and Beagle Brigade (tokens& Cyberdefense Hackathon), Provenance, Marketing Brain, Enterprise Brain. Prepared before the event and finished on site.

Built With

• glasser