{{product.tagline}}
{{product.description}}
-
{{product.name}} reads 13 kinds of logistics document.
Invoices, packing lists, air waybills, bills of lading, customs declarations and more come back as structured JSON.
-
Every field points to the pixels it came from.
Each value carries a page and a box. A reviewer clicks a value and sees the ink it was read from.
-
{{product.name}} runs on infrastructure you control today.
The model runs on one GPU server. [Replace: minimum GPU memory for the 4B model.] Postgres is the job queue, so there is no Redis or Celery.
A scan becomes a checked record in four steps.
Every job passes through these stages. [Replace: current end-to-end time for a single-page invoice on the 4B model.]
The pipeline reads the page.
PaddleOCR runs on CPU and returns text lines with boxes, at about 1.45 seconds a page.
A model extracts the fields.
A vLLM-served model returns JSON that follows a field contract for each document type.
Grounding finds each value on the page.
A matching stage locates each value in the OCR text and records the page and the box.
A reviewer checks the result.
Click a value to see its box on the page. Confidence reflects grounding only, not correctness.
{{product.name}} is a web interface, an API and a pipeline on your own GPU server.
{{product.name}} has three layers. [Replace: the team that builds and runs them.]
Upload a PDF, watch pages and fields arrive, and click any value to see its box. The interface also runs on built-in sample data with no backend.
OCR runs on CPU and a vLLM-served model extracts the fields. A grounding stage then matches each value back to an OCR box.
{{product.name}} runs today on your own infrastructure. PostgreSQL holds the data and the job queue, and one vLLM server holds one model on the GPU.
- React 19
- FastAPI
- PostgreSQL 16
- PaddleOCR
- vLLM
- poppler
People check the result, and the pipeline does the reading.
{{product.name}} reads, extracts and grounds each document. People upload, review and operate it. The chart is an example of a manual document desk. [Replace: confirm against a real customer team.]
The owner decides which document types go live and which mode is the default. [Replace: the owning team.]
Reviewers click a value and check it against the ink on the page.
Submitters upload scans. Operators run the stack, watch the queue and decide when to restart the model server.
The team measures field-F1 on a real 158-document set and re-tests the model as new checkpoints arrive.