How it works
Five stages, and a person at the end of them.
Each stage hands off through a defined interface, so the document reader can be replaced without touching how anything is graded.
The pipeline
Ingest
Load the document alongside the requirements it will be judged against.
Extract
Pull every measured value, whatever the layout, naming, or column order.
Map
Join each value to its requirement from the drawing, spec, or PO.
Grade
Compare against the limits and attach a confidence score and reasoning.
Report
Write the audit trail and the reviewer's copy. Nothing is approved.
In numbers
The reader can be replaced. The grading cannot.
Pulling values off a page is a model's job on a scanned certificate and a plain text extractor's on a native PDF, and the pipeline will use whichever reads the document in front of it. Deciding whether a value meets its limit is deterministic code that never sees the model.
Confidence floor. Below it the grader stops and refers the characteristic to a person.
Requirement forms handled by a single grading engine, across both industries.
Network calls made by the grading engine — checkable in the source.
That seam is why the offline floor still holds: with no model available at all, the pipeline reads what it can with a text extractor and reports every characteristic it could not read as needing review. It degrades into asking a person, never into a guess.
See it run against your own documents.
We will walk through a real review end to end in about twenty minutes — the upload, the grading, the characteristics it hands back, and the sign-off — on a document you choose.