Before
- Open attachment
- Find load and stop data
- Retype every charge
- Spot-check the total
Queue grows with document volume.
The problem
Skilled operators are still hand-typing rate-cons.
$1,299/moroughly, in VA spend
Every loadrekeyed into the TMS
One typocan distort a settlement
Broker PDFs drift: headers move, labels change, scans add noise, and multi-stop loads break assumptions. A generic copy-and-paste workflow turns that drift into manual review or, worse, quiet bad data.
Before / after
After
- Detect broker layout
- Normalize fields
- Validate charges
- Send exceptions to review
Clean records flow; uncertainty stops.
The approach
Adapters absorb layout drift. Confidence gates protect the ledger.
Each broker layout has a small, testable adapter. The output is one stable JSON schema for load IDs, parties, stops, dates, equipment, references, and rates.
Every numeric field must match an unambiguous money token, and the stated total must equal its components. A weak result raises an explicit failure; the pipeline never guesses a dollar amount.
Measured, not claimed
Accuracy generated from the golden evaluation.
| Sample | Adapter | Fields matched | Accuracy |
|---|---|---|---|
| 01_cindercone_classic | cindercone-classic | 25/25 | 100.00% |
| 02_bluekestrel_modern | blue-kestrel-modern | 24/24 | 100.00% |
| 03_moonridge_compact | moonridge-compact | 25/25 | 100.00% |
| 04_cindercone_scanned_noise | cindercone-classic | 24/24 | 100.00% |
| 05_bluekestrel_missing_fields | blue-kestrel-modern | 23/23 | 100.00% |
| 06_moonridge_multistop | moonridge-compact | 36/36 | 100.00% |
| Six-sample total | 157/157 | 100.00% | |
All six PDFs and every entity shown in this case study are synthetic. This fixture score is a regression baseline, not a claim about untested client documents.
About this project
Try it yourself — the whole pipeline is public.
Clone the repository and run make demo to reproduce every number on this page from the six synthetic PDFs. An interactive in-browser version — drop a PDF, watch the extraction, nothing uploaded anywhere — is the next milestone; the build spec lives in the repo's AGENTS.md. Built by Mark Ortega.