FactGuard – Fact-checker for agent output

Verify that a meeting note or promise actually exists in the source before your agent writes it down

Problem

A user found an AI that recorded 'reading club on Friday, time TBD' as 'Friday 19:00'. The format was fine and the save succeeded, but the time was invented. Another builder of a work second-brain stresses that the agent must remember who promised what; a wrong fact becomes a wrong follow-up. Today this is only caught by manual spot checks.

Solution

A lightweight verification service: given the source text and structured facts an agent extracted, judge each fact as supported, unsupported, or contradicted — no rewriting, no decisions. It plugs into meeting-note, CRM-commitment, or second-brain write paths as a configurable hard gate.

Users

Knowledge workers and teams using AI to log meeting notes, customer commitments, or a personal second brain

MVP

CLI and HTTP API entry points; take source plus fact list, output per-fact verdict and confidence; a write hook for one or two note/CRM tools; a human review queue

Revenue

Usage-based API billing plus a plugin subscription; enterprise seats sold as a write gate

Why now

Agents already write more data than humans can audit, and cheap small-model fact checking has been shown workable in the community — a narrow slice indie devs can own