The typical situation
Every incoming raw-material lot comes with a supplier certificate of analysis, as a PDF or a scan. QC staff read each one, check the values against the specification and type them into the lab system. It is slow and it is exactly the kind of work where tired people miss things.
A regulated plant cannot let a model approve anything. It can let a model do the reading, provided a person checks the result and every check is recorded.
What the data looked like
Two years of certificates from about 40 suppliers, the specifications for each material, and the values QC staff typed into the lab system for the same lots, which give the correct answer for every document.
What we built
- Extraction of every test, value and unit from each certificate, with the source location highlighted for the reviewer.
- An automatic comparison against the specification, flagging out-of-limit or missing tests.
- A review screen where QC staff confirm or correct each certificate, with an audit trail of who changed what.
- Accuracy measured per supplier, so a new or messy supplier format is spotted early.
Typical outcome
| Measure | Before | After | Period |
|---|---|---|---|
| Time to process one certificate | about 8 minutes | under 2 minutes | typical outcome |
| Fields extracted correctly on a reviewed sample | typed by hand | about 97% | typical outcome |
What usually trips these projects up
Accuracy is uneven across suppliers. Reporting it per supplier from the first week, not as one average, is what keeps QC staff willing to trust the screen.