Research
Notes from the lab — what we’re building and learning in document AI and vision AI, written for people who ship.
2026-07-21
Document AI
Same encoder, same mistakes: why swapping vision models doesn't fix OCR
We ran three open model families from two vendors, 3B to 14B, against one hard handwritten field. They all failed on the same cards, in the same direction. The reasons say a lot about how VLMs actually read.
2026-06-18
Document AI
Longlichi: a vision-language model for real-world paperwork
Why we build our document stack around a VLM instead of an OCR pipeline — and what that makes possible on handwritten, stamped, and badly photographed paper.