Summary
Every business process has steps and decisions. Robotic process automation has long handled the steps, but the decisions, such as whether a form is signed or a photo supports a refund claim, have usually gone to a person. This video from Sam Witteveen shows how small open image decision models could change that.
The presenter builds a form inspector that takes PDFs or images and runs them through two open models, including image 4B and Omni, to answer yes/no and classification questions straight from the pixels, with no OCR step. Examples include whether a signature is present, whether the date of birth and address are filled in, and what kind of form it is. The models return probabilities for each option, including an explicit “unknown,” so low-confidence cases can be routed to another model or a human reviewer. Settings such as image DPI are also explored, along with some of the issues that come up in practice.
The video also looks at why traditional RPA vendors like UiPath struggled. UiPath went public in 2021 at a valuation of around $30-40 billion, and its market cap is now under $6.5 billion, a decline Witteveen attributes to models that needed heavy per-use-case fine-tuning. The top model on the new image benchmark was reportedly built by one person in 15 days using $1,200 of rented GPUs.
📺 Source: Sam Witteveen · Published October 02, 2026
🏷️ Format: Hands On Build
