Quote Parser: nine supplier formats in, clean line items out, with an analyst checking every row.

Healthcare organizations buy equipment from suppliers like GE Healthcare, Philips, and Siemens, each pricing quotes differently. Before an analyst could apply any judgment, they had to retype every line into a structured format: PDFs, scans, spreadsheets and emails, in at least nine distinct layouts. Line items per quote were climbing every year, and the team preparing them was getting smaller.

Pricing analysts, about 10.5 FTE in 2026, own the final number the customer sees, so any tool had to leave them in control rather than replace their call.

The analyst uploads a quote. The model reads document context, not fixed coordinates, and returns every line item: product, description, quantity, list price, quoted price. Where a price can be derived from what the supplier gave (List = Net ÷ (1 − Discount)), a deterministic rule computes it. Where a field genuinely can't be derived, it comes back as "N/A" instead of a guess. That's the rule I care about most: a blank cell gets checked; a wrong number that looks right gets shipped to a hospital.

The architecture stayed boring on purpose: one prompt, no agents, no retrieval, SQLite and lightweight auth. Claude runs on AWS Bedrock because quote data stays inside the company's own cloud account, which is what got this past security review.

This extraction engine is one slice of a larger legacy-modernization effort, Phoenix. It shipped first, standalone, because it proved the AI worked on real quotes before the platform rebuild absorbed it. More on Phoenix →

Solve the right problem. The bottleneck wasn't document extraction. It was everything analysts did before benchmarking could even start.

Build the labeled test set on day one. I know the tool works because analysts say so and the scorecard moved, but I can't prove extraction accuracy with a number. Phoenix fixed this with an evaluation harness.

The best metric was analyst-reported. "I can breathe now," from one of the leads, mattered more than the throughput number.