Built for the whole batch,not the perfect demo.
Reading, structure, job state, and review stay connected.
The model reads. The system carries the result.
Page pixels become structured data without losing ownership or review context.
Four layers make the output usable.
Preparation, structure recovery, and durable batch state make extraction reviewable.
- 01
Vision pass
Rotation, contrast, patches, and page-shape preparation turn a document page into visual tokens.
- 02
Document prompt
A 7B-class Qwen2-VL and olmOCR-style path is guided toward handwriting, forms, and tables.
- 03
Cell graph
Headers, rows, totals, merged areas, and column relationships are rebuilt before export.
- 04
Owned job flow
Queue state, file metadata, retries, and downloads surround the model so a real batch can keep moving.
A workbook asks more than plain OCR.
Headers, rows, totals, and merged cells stay connected in one schema.
One request starts the work. A durable job finishes it.
Queue state, workers, file ownership, and retry-safe output stay together until review.
Queue
The run is admitted and tracked outside one web request.
Worker
Pages are prepared, read, classified, and structured.
Storage
File ownership, result metadata, and retry-safe output stay together.
Review
Source, cells, corrections, and downloads remain connected.
The endpoint is a spreadsheet a person can trust and change.
Source and table stay side by side for corrections, reviewed status, and clean batch downloads.