Computer Vision

Our computer vision work covers recognition, detection, segmentation, image generation, and multimodal systems that combine vision with language. Vision language models are part of that picture but not the whole of it, and a good deal of what we do sits in ordinary supervised vision, where plenty of the hard problems are still open.

Much of the effort goes into data and evaluation rather than architecture. Benchmarks reward average accuracy, which hides the handful of cases a model gets badly wrong, and curated training sets can quietly lose the properties that made the original data worth having. We build benchmarks and auditing methods that surface those problems before a model is trained on the data, or deployed on the strength of a leaderboard number.