Technology
The technology behind better document workflows
Our work combines recognition models, our own font and glyph data, document-processing methods and structure-aware editing.
From data to products
- DataFonts, character fingerprints, checked glyphs
- RecognitionCharacters and text lines · AURORA OCR
- StructureLayout, sections, fields · AURORA Resuminer
- EditingChange content, keep the design · AURORA PDFit, RESUMagic
- ProductsTools for people and technology for teams
Each product uses the technologies its task needs. The diagram shows our areas of expertise, not one model shared by every product.
Character and font intelligence
Every recognition and editing task starts with the shape of a character in a particular font. We collected and checked this data ourselves.
Recognition
A character model built on ConvNeXt-Small (≈ 49.5 M parameters) reads single characters at ≈ 99.85 % on clean and ≈ 99.65 % on scan-like characters in our internal test. The next stage reads whole text lines; preliminary line accuracy is about 93–94 %.
Document structure
A document is more than text: columns, headings, blocks and the relations between them. For résumés we map this structure to a schema of 234 fields in 21 blocks.
Processing and editing
Editing a PDF means changing content while the typography, layout and graphics stay as they were. PDFit uses the fonts embedded in the document; RESUMagic, in development, adds an understanding of résumé sections and blocks.
Training and evaluation
Models are trained with hard-example mining, targeted retraining and simulated scanner and camera damage. They are tested on held-out internal data and on public benchmarks, olmOCR-bench and UNLV.
Want to use this technology?
Tell us about your documents and what you need to achieve.