Warrant — a support AI that can't act without proof
Link to open source: https://github.com/giri-harsh/warrant
When a customer email arrives, it extracts the relevant facts, retrieves supporting evidence from a real knowledge base, and only drafts a reply once it has evidence for every part of the question. If it can't find enough evidence, it says so instead of guessing — a wrong answer with confidence is worse than no answer at all.
Every outbound action — sending a reply, filing a ticket — is blocked by default at the platform level, not by application logic. A human must explicitly approve before anything can send. This isn't a setting the app chooses to enforce; it's structurally impossible to bypass, because the underlying platform (Swytchcode) rejects any action without a valid approval token, regardless of what code calls it.
To prove this isn't just a claim, Warrant keeps its own action ledger and independently reconciles it against the platform's own audit log — a record the app itself doesn't control. During testing, we discovered the approval check only verified a token existed, not that it was valid, meaning a forged token could bypass the policy gate. The reconciler still caught it: the forged action reached the provider, but with no matching warrant behind it, it was automatically flagged as off-book. That's the core idea in action — not "trust the AI," but two independent records that can't both be wrong.
The full pipeline — extraction, retrieval, confidence scoring, escalation/dedupe routing, the approval gate, and audit reconciliation — is live and verified end-to-end, including a real round-trip through Gmail and Resend.
