Image Decision Models for RPA: Forms, Scans and Screenshots

Sam Witteveen returns to decision models in this 16-minute video, this time for images, with robotic process automation (RPA) as the use case. The description links two open models on Hugging Face: ImaJev 4B and Jev Omni.

The opening chapters set out the problem. At 0:17 he makes the point that RPA automates steps, not decisions. At 0:39 he covers why custom deep learning models fell short for those decisions, and at 0:59 he introduces the Form Inspector app and the two models. He explains the focus on images at 1:38, and at 1:56 describes the job itself: checking screenshots, photos and forms.

A short history follows. The RPA story starts at 2:42, and what went wrong with RPA takes the chapter at 3:24. At 5:05 he looks at the ImageBench leaderboard, then at 5:24 tells the story behind ImaJev. The idea of smart if statements with confidence scores comes up at 7:29.

The demo of the Form Decision Inspector starts at 8:01. From 9:11 he runs ImaJev 4B against Jev Omni. At 10:53 he compares choice questions with true/false questions, and at 11:57 he uses the answers to drive follow-up emails with conditional logic. Adding new questions is at 12:41 and a second form at 13:21. The last chapter, from 14:35, sets confidence thresholds per question.

The model cards add some detail. ImaJev 4B is a LoRA adapter on Qwen3.5-4B, released under Apache-2.0, and it returns a probability for each option, including an explicit "can't tell". Jev Omni is built on Gemma 4 12B IT, is also Apache-2.0, and accepts text, images, audio and video.