A web app that reads valve specification PDFs and pulls out structured, quotable data automatically. Built for the oil and gas sector.
Every valve quote starts with a multi-page specification document. Over 50 data points per document, previously extracted by hand.
The annual cost of manual specification extraction across the engineering team, before automation.
Pressure ratings, temperature ranges, material grades, certification requirements, and dozens more fields per document.
In oil and gas, a misread pressure class or skipped corrosion testing requirement has real consequences downstream.
Getting the prompt right took the longest. We built a schema covering the full valve specification taxonomy: operational parameters, physical specs, materials, design standards, testing requirements, operator-specific fields. The first few attempts returned patchy results. Fields missing, values in the wrong format, nested data flattened into useless strings. It took several rounds of iteration against real spec documents before the extraction was consistent enough that the client’s engineers actually trusted it over doing the work themselves. That trust threshold was the real milestone, not any particular technical achievement.
Valve specs can also be enormous. Some documents are complex enough that the AI hits its response limit before finishing the extraction. We could have truncated the output and accepted the loss, but you cannot do that with safety-critical equipment specs. So the system detects when a response has been cut short, automatically continues the conversation to get the rest, and stitches it together. No fields lost, regardless of document length.
Then there is the format problem. Every valve manufacturer’s datasheet looks different. Different fonts, different structures, different ways of presenting identical information. The parsing layer has to handle all of it, because nobody at the client has time to manually prep documents before uploading them.
We also had to deal with processing time. AI extraction is not instant, especially on longer documents. The system processes everything asynchronously, so an engineer can upload five PDFs and go do something else. The dashboard shows results as they come in. It sounds like a small thing, but if the UI had blocked during processing, nobody would have used it.
The structured output goes straight into the quoting workflow. No retyping. No transcription errors on material certifications or compliance standards.
6 mins
What used to be a painstaking manual process, often taking hours per document, now completes in under seven minutes.
97.3%
Across 120 test documents, the system correctly extracted 97.3% of fields without manual correction.
200 days
Over 200 person-days of engineering time recovered each month, redirected from data entry to actual engineering work.
“We used to dread spec sheets. A senior engineer would spend half their morning on data entry before they could even start quoting. Now the system handles it in minutes and we barely think about it.”
Senior Engineer, UK valve supplier
What actually surprised us was how much the consistency mattered to them. More than the speed, even. An AI does not skip fields at 4pm on a Friday. It does not misread a pressure rating because the font is small on page seven. It picks up the fiddly nested stuff that manual processing tended to miss: temperature ranges with min and max values, multi-standard certification requirements, corrosion testing thresholds with specific ferrite content limits. Extracted the same way, every time.