A 20-year-old, 366-screen legacy monster, rebuilt as a modern application.

For roughly 20 years, a spend-analysis team ran their entire quote-benchmarking business on a PowerBuilder desktop app over a VDI session. The core daily job was manual: a vendor quote arrives as a PDF or scan, an analyst reads it and types every line item into a form by hand, then compares prices against market benchmarks. That transcription step was pure data entry, and it came before any of the analyst's actual judgment.

I reverse-engineered the legacy system's scope directly from its codebase: 21 major modules, 366+ screens, a genuinely enormous surface area. But the single biggest pain point was narrow: manual line-item transcription, the one step that was pure mechanical cost with zero judgment value. Five-whys on "why is this still manual?" landed on the real answer: the assumption that a human must transcribe quotes was correct for 20 years, right up until vision-capable LLMs made it tractable, and nobody had re-checked it since.

Rather than porting the legacy app screen-for-screen, a multi-year program that would have faithfully reproduced the same manual bottleneck, I rebuilt it as a modern web application: roughly 60–70% of the legacy system rebuilt completely, prioritized by what analysts actually use daily, plus the AI extraction capability the legacy app never had.

The backend started on Python (FastAPI). Partway through the build I learned the rest of the company's applications run on .NET, so I rebuilt the backend on .NET 8 to match the enterprise stack. That is the backend in use today; the Python version is retired.

Benchmarking works differently depending on what's being priced: capital equipment is compared by discount from list price, consumables by price percentile across other hospitals. A tornado chart shows which lines drive the gap, and reports export in two formats.

The extraction capability proven inside this prototype shipped separately as a standalone production tool, AI Quote Automation, now in daily analyst use.

Quote Parser, the production tool, shipped without a labeled accuracy number. Phoenix built that harness: every extraction scored against hand-labeled ground truth for recall, precision, and whether each row is perfect. The prompt grew from about 25 lines to about 180 as the harness exposed failure cases.

The hardest case was a hospital bed quote with subtotal rows that look exactly like products. The model has to find the 10 real line items and ignore the 3 subtotals. It does.

A prioritized gap analysis is the real scoping weapon. It turns "the legacy app is huge and scary" into a rational, defensible plan instead of an excuse to sprawl.

Re-examine old assumptions when the underlying technology shifts. The 20-year manual step wasn't a bad decision. It was correct until it wasn't, and nobody had re-checked.

Grow the evaluation set before quoting an accuracy number. 100% on 55 line items across 2 quotes proves the harness works, not that the model is perfect.