A commercial contractor's bid team gets quote requests by email: drawings attached or linked, a date buried somewhere, an engineer's seal on a sheet. Someone used to open every one, work out whether it was even a request, dig out the facts and key them into the bid list. This pipeline does the reading and the typing. A language model decides whether an email asks for pricing and reads the key facts off the drawings; plain code does everything else, from scanning attachments to checking the engineer against the state roster and writing the row.
The demo is the real flow with invented emails and a scripted trace, so you can see what each stage does and why a row lands where it does. Anything the pipeline is not sure about goes to a review card instead of the bid list, and a person's reaction decides it. The thinking behind the build is in the blog post Six things I learned building an AI quoting workflow. Confidences and timings in the demo are illustrative.
A short, free conversation about what's working, what isn't and whether we can help. No pitch, no pressure.