I enjoy building infrastructure that removes friction for other engineers. Open source is also a useful test: an idea has to make sense outside the team and context where it began.
The projects below try to make capable ML systems easier to build with, extend, and understand.
Ultralytics became the world's most popular open-source computer vision framework, with more than 60,000 GitHub stars. I co-authored the YOLOv8 model family and worked across YOLOv5 and the framework on classification, segmentation, experiment logging, distributed training, and Explorer, a semantic and SQL interface for understanding computer vision datasets.
An integration is useful when it removes a decision and a piece of glue code. At W&B, I made experiment tracking native to widely used ML libraries: the sweeps backend in Hugging Face Transformers, integrations for Catalyst, Tianshou, and CycleGAN/pix2pix, and Sacred support in the wandb client.
Retrieval systems should make it possible to combine semantic, lexical, and late-interaction methods without assembling a new infrastructure stack. In LanceDB, I worked on rerankers, hybrid search, and multivector and ColPali support, then brought those capabilities into LlamaIndex, LangChain.js, the Retriever API in AutoGen, Letta, Agno, FiftyOne, and multimodal-QA support in Meta's synthetic-data-kit.
stable-worldmodel exists so researchers can compare ideas on a common platform instead of rebuilding the surrounding system for every experiment. I built its Lance-based video data layer, streaming dataset writers, and dataloader benchmarks.