Six weeks of testing on their own drawings settled a question a century-old engineering firm had worked on for a year.

UCC Environmental has been designing material handling and emissions control systems for power generation and heavy industry since 1920, and has more recently expanded into water treatment through acquisition. Roughly 300 people work there and about 70% of them work directly with mechanical engineering drawings, which makes the company’s accumulated design history one of its principal working assets. Competing in price-sensitive adjacent markets such as mining and cement puts pressure on both the cost and the turnaround of a proposal, and each of those depends on how quickly an engineer can find what the company has already built.

That history amounts to roughly half a million drawings held in SharePoint and searchable only by title and drawing number, with a numbering convention that had changed several times across the decades and was due to change again. Systems engineering produces six to seven contract drawings and proposals a week and needs to search on parameters that appear nowhere in a title block, such as storage capacity, pressure rating, or metallurgy, while parts teams field customer calls about legacy equipment and support teams investigate field problems against drawings that may predate the current ERP entirely. When a drawing can’t be found, the engineer has to design it from scratch, and that can cost up to two weeks. UCC had spent roughly a year working the problem internally, scanning drawings into searchable PDFs and building a Microsoft Copilot agent on top of them, and had run into a file-size processing ceiling and answers inconsistent enough that people had begun to stop trusting the tool.

We built the six-week engagement around a head-to-head extraction test on UCC’s own drawings. The market for AI drawing extraction is full of purpose-built tools with convincing demos, and every demo runs on drawings the vendor picked, so we wanted to know what a century of inconsistent source material would do to them. Our working assumption going in was that a specialist vendor would win.

We spent the early weeks in Chicago interviewing across systems engineering, support and warranty, parts marketing, and the general product division, auditing the library itself, and meeting seven vendors spanning extraction-only, search-only, and integrated platforms. For the test we defined the metadata schema we wanted to search against, manually extracted the correct values from a sample of drawings to establish ground truth, asked UCC to dig out their oldest hand-lettered originals so that the sample wouldn’t flatter anyone, and scored each vendor’s API output field by field. The recommendation was delivered as a single document covering the architecture, a three-release feature roadmap, Azure infrastructure specifications, a monthly operating cost model, and a costed fourteen-week implementation plan, written so that someone with no prior context could pick it up and build from it.

Across the test set, Infrrd, the specialist vendor we had expected to recommend, averaged 51% accuracy on field extraction while Gemini averaged 97%, returning perfect extractions on most of the drawings, including sheets hand-lettered in the 1960s. Infrrd requires its model to be trained on each drawing variation and its target fields to be defined in advance, which works for a standardized library and struggles with one assembled over a hundred years, and Gemini did the same job with no training at a per-drawing cost orders of magnitude lower. That result moved the recommendation to a modular architecture, Gemini for extraction and Elasticsearch for hybrid search, running in the low four figures a year against the hundreds of thousands quoted by the integrated enterprise platform we had evaluated alongside it. Both client sponsors adopted the recommendation formally at the March readout.

Our impact model estimated six figures a year in recovered engineering time, with several times that again in rework avoided on drawings that exist somewhere in the archive but can’t be located. Separating extraction from search also leaves UCC with a structured record of half a million drawings that can be queried outside the search tool itself.