Upload dealer claim PDFs. The pipeline classifies each one, routes it to the right
extraction tier, pulls what it can, and runs every value through the validation gates. It shows
you what it got, how confident it is, and exactly where it refused.
Engine
No files selected
Two pipelines, deliberately.Textract + geometry reads the page as text with
coordinates, then locates each field using label patterns and spatial rules. VLM layout inference
looks at the page and returns the fields directly, reasoning about what a value sits under or beside.
The difference matters because the geometric path generalises only as far as its label list and its
pixel windows reach. Tuned against three documents it fits three documents: tightening the windows
recovered one invoice's odometer and lost another's totals, and loosening them did the reverse. The
layout model needs no per-vendor patterns at all, which is the argument for it.
What is universal in either pipeline is the validation. The VIN check digit, consensus across repeated
readings, format constraints and arithmetic reconciliation all hold without knowing anything about the
layout. Compare both runs the two side by side and separates confident failures from values that
were present on the page and missed.
Handle with care. Uploads are processed in memory and deleted immediately; nothing is written to
storage and no document content is logged. Even so, this endpoint is public. Do not upload documents
containing real customer data that you would not be comfortable sending over the internet.