Orders, deliveries and invoices. One connected review.
Real demo recording · English narration and captions · Fictional data
THE PROBLEM
Differences between ordered, delivered and invoiced items can be missed when each document is reviewed separately.
THE WORKFLOW
Extract fields from three fictional documents using local Qwen and LightOnOCR for scans. Check quantities, prices and totals with exact decimal arithmetic, retain the originals and save reviewed corrections.
01Three-document intake
02Local extraction and OCR
03Matching and human review
04Revisions and export
RECORDED EVIDENCE
Results, with context.
Results on synthetic data
Check
Cases
Bundles: all fields correct
3/3
Fields matching the reference
171/171
Correct reconciliation outcome
3/3
Dataset
original-controlled-fixtures-v1
Completed cases
3/3
Median / p95 per case
158.648 s / 162.037 s
Evaluation record (UTC)
2026-10-08
Source: the project’s evaluation record, using local inference. Recorded hardware: NVIDIA GeForce RTX 4070 Laptop GPU / 8 GB. Latency includes the full HTTP case and may include initial model loading. These measurements have not been rerun on the machine serving this website.
A document reconciliation pilot with visible sources, exception rules and structured exports.
What we would define first
Order, delivery and invoice formats, line identity, tax rules, acceptable discrepancies and the destination system.
Demonstration limits
Three controlled bundles: matched, discrepant and scanned. OCR and extraction require review; absent values stay absent. Financial approval is not automated.
Evaluated demo