From faxed invoices to a queue you can trust.
A restoration contractor was re-typing supplier invoices into claim files by hand. We built an extraction pipeline that reads the documents, proposes the line items, and asks a human only when it is unsure.
- Client
- Coastline Restorations
- Industry
- Insurance restoration
- Location
- Calgary, Alberta
- Timeline
- 12 weeks
- Team
- 2 engineers, 1 designer
Intake queue
Extracted fields
94% confidenceClaim Intake & Line-Item Extraction
Every claim arrived as a folder of PDFs: supplier invoices, adjuster notes, photos, and a hand-typed estimate. A coordinator rebuilt the line items by hand before the file could move.
Volume doubled in storm season, and the errors that crept in during the rush were expensive to unwind after the fact.
- Python
- OpenAI
- PostgreSQL
- S3
What it looks like in use.
Extraction audit
4 rowsManual review load
BeforeAfterFour decisions that made it work.
-
01
Read the worst documents first
We tuned against 600 real invoices — including scans, faxes and three suppliers who still use dot-matrix.
-
02
Never guess silently
Below a confidence threshold, the field is left blank and highlighted for review rather than filled with a plausible number.
-
03
One screen for review
The coordinator sees the document and the extracted fields side by side, with a single keystroke to accept the page.
-
04
Learn from corrections
Every correction is logged with its supplier and document type, and feeds the next monthly tuning pass.
“It does the boring part perfectly and it tells us when it is unsure. That is the whole reason the team trusts it.”