Getting a company's proposal library out of a tool that had no export

A sales organisation ran a large pipeline where every opportunity linked to a proposal held in a third-party document tool. When a deal closed, a billing record was created downstream that needed the actual proposal document so it could be read automatically and turned into line items.

The document tool offered no export path for this. So proposals were being pulled out by hand, one at a time, on every closed deal — a manual step sitting directly in the revenue process.

We built an automated mirror. The system walks every workspace in the document tool, retrieves each proposal in both formats, and mirrors it into the company's own storage and billing tracker — triggered automatically when a deal closes, and runnable on demand from an admin console with live log tailing.

A second phase added generation as well as retrieval: a document-builder agent that assembles new proposals from reference material and brand tokens rather than having someone copy an old one. It is multimodal — it reads images attached to a request within enforced size and count bounds — and returns structured proposal data validated against a schema, with an automatic repair retry when the model returns something malformed. Build jobs queue in the database and are drained by a background worker, so a slow generation never blocks the interface.

A small detail worth noting for anyone shipping commercial software: the PDF library was chosen specifically to avoid copyleft licensing obligations on server-side use. Dependency licences were treated as an architectural constraint, not an afterthought.