Apr to May 2026
AI in the pipeline
Vision and LLM steps that fail safe to a person.
The problem
About a quarter of uploaded documents needed someone to crop them by hand, and free-text inputs blocked checks from starting at all.
What I built
- A YOLO and OpenCV service that finds the document in a photo, straightens it and closes the ops task when it is sure.
- LLM classification of free-text input that fails closed: anything ambiguous still goes to a person.
- Reporting-manager responses that go straight to verified and QC instead of waiting in an intermediate state.
- Stuck checks that route themselves to research instead of raising Slack alerts.
Stack: Python, YOLO, OpenCV, LLM.
What changed
- 19.5% to 43.8%
- correct on 411 production images, 2.25x recall
- Fails closed
- ambiguous cases still go to a person
Source: Offline benchmark on 411 production images. The production-level effect was not measured.
Your version: Someone checks documents by hand?
Vision and LLM steps that fail safe to a person instead of guessing.
Tell me about itNext case: Pawse