Somehow ended up owning the computer vision side of things here, further than I expected to go. It began with training object-detection models and grew from there, later touching some llm work as well.
On the vision side, I trained and iterated on computer-vision object-detection models. That meant sourcing and versioning image datasets, then tuning preprocessing and augmentation strategies across many training runs, each pass chipping away at the error rate until the detection accuracy was reliable enough to ship.
From there I built the end-to-end real-time computer-vision pipeline: object detection and tracking, event recording, and automated report generation. Rather than staying a research notebook, the whole thing was wired into a web application, so the models run against real footage and turn what they see into structured reports.
Working on it...