Teaching a Machine to Read a Quote Request Like an Estimator Does

The challenge

Customers sent requests for quote as spreadsheets and emails, every one in a different layout, using their own abbreviations for the same parts. Each line had to be matched by hand against a catalogue of nearly twelve thousand items spanning dozens of product categories — work that needs real domain knowledge of material specifications and schedule-to-wall-thickness conversions, and that only a handful of people could do.

Where we came in

The tempting framing is a parser: teach software the formats and it will read them. That fails on contact, because the variation is not in the layout but in the language. We treated it as a matching problem instead — represent the catalogue by meaning rather than by string, and let the system propose candidates for a human to confirm.

What we did

• Ran a discovery and architecture phase before writing production code, delivered as a written report rather than a verbal readout
• Built a data engineering and embedding pipeline that represents the product catalogue by meaning rather than exact text
• Developed a parsing and matching engine with a web interface for reviewing and confirming proposed matches
• Stood up a dedicated evaluation phase so accuracy is measured against a held-out set rather than asserted
• Delivered in numbered phases with a demo closing each one

Where it landed

• Inbound quote requests are parsed and matched against the catalogue rather than transcribed line by line
• Matching accuracy is measured against a validation set, so changes to the engine can be judged rather than guessed at
• Delivered through discovery, data engineering, matching, and evaluation phases to handoff