Inspiration
coding agents got trusted when the workflow stopped depending on the model being right. Finance agents have none of that.
What it does
matcher clears 17 of 29 lines; 12 exceptions become isolated cases; four read-only tools; structured proposal → Finance PR → 18 deterministic checks → blocked proposals go back to the same worker with the engine's own text → controller merges → hash-linked sandbox ledger → close. Plus the Control PR, the plain-English composer, and evidence upload.
How we built it
three lanes, one contract; the sponsor integrations described as what they actually are.
Challenges we ran into
the model was too good for our demo; we deleted a flattering bug; a control was silently emptying the Auto lane; our safety headline was wrong; parallel builders drifted; deploying found what local runs never would.
Accomplishments that we're proud of
the three-run table above, plus: the system refuses well, and nothing on screen overclaims.
What we learned
running it for real is a different activity from building it; report the distribution, not your best run; a fingerprint has to be a promise you keep.
What's next for Verity
the practitioner review (pack is ready and needs an hour), more of the rule surface, durable state, and the workflows we deliberately did not build.
Built With
- ao
- dodopayments
- maximor
- typescript
Log in or sign up for Devpost to join the conversation.