Research areas

Overview

Research areas

Document understanding

Contracts, invoices, forms and reports carry information that systems need as structured data. We build extraction pipelines that read real-world documents, including scans and complex layouts, and know when to ask a person instead of guessing.

  1. document

    PDF · scan · photo

  2. read

    OCR · layout

  3. extract

    fields · tables

  4. validate

    rules · confidence

  5. data

    JSON · your system

Fig. 1A document pipeline that routes uncertain fields to a person instead of guessing

Typical problems

  • Templates and rules break every time a supplier changes their layout
  • Scanned and photographed documents are too noisy for plain OCR
  • Extracted values look plausible but are silently wrong
  • Every document still needs manual checking, so automation saves little time

What you get

  • An extraction pipeline for your document types and layouts
  • Field-level confidence scores and validation rules
  • A review queue for the cases that need a person
  • Accuracy measured per field on a labelled test set

How we approach it

We start with a labelled sample of your real documents, including the bad scans and unusual layouts, and measure accuracy per field rather than per document. That shows exactly where the system can be trusted and where it can’t.

Each extracted value carries a confidence score and is checked against business rules, like totals that must add up or dates that must be in range. Anything uncertain goes to a person, so the automation handles the routine cases and people handle the hard ones.