Document extraction is one of those practical AI challenges that rarely gets the spotlight but makes a huge difference in real-world automation. Landing AI, the company chaired by Andrew Ng, has announced an update that directly addresses this issue: the Document Pre-trained Transformer 2 (DPT-2) now powering what it calls Agentic Document Extraction (ADE).

In many business workflows, scanned invoices, unstructured tables, and signatures embedded in PDFs can easily confuse traditional extraction systems. These systems often lose the document’s layout and, with it, the meaning of the data. Landing AI’s new model is designed to keep that structure intact.

According to the company, DPT-2 introduces:

1. Cell-by-cell parsing of complex tables, maintaining alignment and context.
2. Improved layout detection that works even on low-quality or skewed scans.
3. Broader support for visual elements, including checkboxes, QR codes, and handwritten signatures.
4. The model’s ability to “ground” data meaning it links extracted text to its original position in the document could make downstream validation much easier, especially in industries like finance and logistics.

Andrew Ng shared a walkthrough demonstrating how ADE with DPT-2 processes complicated layouts without losing accuracy, a step forward for organizations dealing with high document variability.

While it’s still early to see how DPT-2 performs at scale, this release signals a clear push toward more reliable document AI one that doesn’t break the moment things get messy.