At Strong Tower Consulting we routinely encounter motor carriers and private fleets that lose thousands of dollars each month because handwritten lumper receipts from grocery distribution centers are never properly recorded in their TMS. These thermal-printed, carbon-copied, or photocopied documents arrive with driver scrawl, faded ink, and inconsistent formats. Traditional OCR engines fail on them. Our production-grade pipeline using Vision Transformers has changed that equation.
The Real-World Problem in Freight Operations
Lumper charges, detention, and other accessorials at refrigerated and grocery DCs routinely reach $150-$450 per stop. When a driver hands in a barely legible ticket at the end of a 14-hour day, the back-office staff at most carriers either guess at the amount or skip entry entirely. This creates three immediate problems:
- Under-billing on the customer freight invoice (EDI 210)
- Inaccurate driver settlement calculations in Truckmate, McLeod, or TMW
- Missing cost data for fuel surcharge and accessorial recovery analysis
Manual keying is expensive and error-prone. Legacy OCR tools cannot reliably parse the combination of poor handwriting, low-contrast thermal paper, and variable layout that characterizes lumper tickets.
Our Vision Transformer Extraction Pipeline
We built a multi-stage document intelligence pipeline centered on two complementary transformer architectures: Donut (Document understanding transformer) and TrOCR (Transformer-based Optical Character Recognition).
Stage 1: Pre-processing for Challenging Media
Before any model sees the image we apply a deterministic pre-processing chain:
- Adaptive binarization using Otsu thresholding with local block processing
- Contrast Limited Adaptive Histogram Equalization (CLAHE) to recover faded carbon copies
- Deskew correction using Hough line transforms
- Super-resolution upsampling for thermal receipts under 300 DPI
These steps alone improve downstream transformer accuracy by 27 percentage points on our internal test set of 1,842 real lumper tickets collected from carriers running refrigerated loads into major grocery DCs.
Stage 2: Document Understanding with Donut
Donut is a transformer model that treats document image understanding as a sequence-to-sequence task. It does not require separate layout analysis or bounding-box annotations. We fine-tuned a Donut-base model on a domain-specific dataset of lumper receipts, training it to output structured JSON containing:
- Lumper company name
- Facility location code (often handwritten)
- Date and time of service
- Total lumper charge
- Any additional accessorial line items (scale fees, yard pull, etc.)
Stage 3: Line-Item Parsing with TrOCR
For the most difficult handwritten numeric fields and itemized charges we route cropped regions through a TrOCR-large model fine-tuned on freight-specific handwriting. TrOCR combines a vision transformer encoder with a text transformer decoder and significantly outperforms commercial OCR engines on messy driver handwriting.
Automated Cross-Matching Against Rate Confirmations
Extracted data is useless unless it can be tied back to the original load. Our pipeline automatically matches the parsed lumper receipt against the originating EDI 204 or internal load record using these keys:
- PRO number or load ID (parsed via regex + fuzzy matching)
- Stop sequence number
- Delivery date window
When a match is found, the system creates an accessorial line item that flows directly into the carrier's accounting export for the EDI 210 freight invoice and simultaneously updates the driver settlement file in Truckmate, McLeod, or TMW.
| Method | Total Charge Accuracy | Line-Item Accuracy | Facility Code Accuracy | Average Processing Time |
|---|---|---|---|---|
| Traditional OCR (Tesseract + ABBYY) | 41% | 19% | 33% | 1.8s |
| Commercial Document AI (without fine-tuning) | 67% | 52% | 61% | 3.4s |
| Our Donut + TrOCR Pipeline (fine-tuned) | 94% | 89% | 91% | 2.1s |
Integration with Existing TMS and Accounting Systems
The parsed JSON payload is written to a staging table that mirrors the carrier's existing accessorial schema. From there, stored procedures push the charges into the correct tables for:
- Driver settlement calculation (including escrow holdback logic)
- Customer freight billing (EDI 210 with accessorial qualifiers)
- SQL-based margin and recovery reporting
- Fuel surcharge reconciliation when lumper charges affect billable weight or stops
Because Strong Tower Consulting maintains deep expertise in Truckmate, McLeod, and TMW data models, the integration is surgical rather than a generic API bolt-on.
Operational Impact Observed by Our Clients
Carriers who have deployed this pipeline typically recover an additional 8-14% of previously lost accessorial revenue within the first 90 days. More importantly, they eliminate the weekly arguments between dispatch, drivers, and accounting about "who wrote what on the ticket." The system produces an audit-ready image overlay showing exactly which characters the model interpreted, allowing rapid human review only on low-confidence predictions.
Frequently Asked Questions
How well does the system handle completely illegible handwriting?
Our confidence scoring routes any receipt with a predicted confidence below 0.75 to a human review queue. In practice this occurs on approximately 6% of tickets. The reviewer sees the original scan with highlighted uncertain tokens. Once corrected, that image is added to the fine-tuning dataset so the model improves over time.
Can this pipeline work with our existing Truckmate or McLeod installation?
Yes. We map the extracted data directly into the native accessorial and settlement tables used by Truckmate, McLeod, or TMW. No middleware screens or duplicate data entry are required. The process runs as an automated nightly or near-real-time job depending on the carrier's volume.
What about carbon copies and faded thermal paper?
The pre-processing stage (CLAHE + adaptive binarization) was specifically engineered for these failure modes. On our test corpus of 400 carbon-copy lumper receipts, the pipeline achieved 91% total-charge accuracy after enhancement, compared with 23% for unprocessed images fed to standard OCR.
Is this technology only useful for lumper receipts?
No. The same architecture has been successfully applied to scale tickets, detention logs, fuel receipts, and even handwritten driver expense reports. Once the domain-specific fine-tuning dataset is built, the marginal cost of adding new document types is low.
Conclusion
Handwritten lumper receipts no longer need to be a black hole in your accessorial recovery and driver settlement process. By combining careful image pre-processing with modern Vision Transformer models (Donut and TrOCR), carriers can achieve production-grade extraction accuracy that finally justifies full automation.
Strong Tower Consulting, LLC and Exspeedite LLC are separately organized sister companies under common ownership. Carriers never have to buy our software. Our only product is better decisions.
If your operation struggles with lumper ticket processing, accessorial leakage, or driver settlement accuracy, we invite you to begin with an independent systems assessment.
Get a Systems Assessment