How AI remediation works
What our pipeline does behind the scenes.
Updated
Our remediation pipeline runs in four stages.
1. Extraction
We parse the source PDF using a structure-aware extractor (not just text). We pull out images, layout regions, tables, form fields, and any existing tags.
2. Structural reconstruction
A model identifies headings, paragraphs, lists, and tables based on visual and textual cues. We rebuild the document's logical structure tree.
3. AI annotations
For each image, we generate alt text using a vision model. For each table, we identify header rows and columns. For form fields, we infer labels.
4. Re-encoding
We rewrite the PDF with proper Tagged PDF structure, language tags, and a navigable bookmark tree. The visual layout is preserved bit-for-bit.
Where humans come in
The "Edit" step lets you review and correct AI suggestions before publishing. We strongly recommend it for public-facing documents — AI alt text is good but not perfect, especially for charts and diagrams.
Privacy
Documents are processed in our infrastructure. We don't share content with third parties. After 30 days of inactivity, both source and remediated copies are deleted from our storage automatically.
Was this article helpful?
Your feedback helps us improve our docs.