Trase began with a narrow promise: upload a bank export, see which receipts are missing, and let an agent retrieve them.
What worked
The system had to do more than click. It identified vendors from transactions, worked through their portals, connected the retrieved document back to the transaction, and kept the result in one place. The architecture separated judgement from action: edge functions decided what to do; a browser extension did the work in the portal.
Bank export → identify vendor → vendor portal → retrieve document → match transaction → human review
Why it stopped
The feature was useful and the work was substantial. But retrieving a missing receipt was too narrow a problem to sustain a company on its own.
I stopped the product rather than stretch a working feature into a company.
What remains
The part worth keeping is what it revealed: a professional record is not only a set of files. It is a chain of evidence, uncertainty and responsibility. The difficult design question was never whether an agent could make a click. It was how to make each hand-off legible enough that someone could trust the result. That design question remains useful beyond Trase.