I enjoy building products around difficult machine learning problems.
Most of my work lives at the intersection of research and engineering: exploring new ideas, building the systems that make them practical, and figuring out how they reach people. Right now, I'm working on retrieval and multimodal AI infrastructure while independently researching representation learning, world models, and alternative architectures for intelligence.
I believe the most meaningful research doesn't end with a paper. It becomes something people can build with.
Outside ML, I read philosophy and listen to a lot of music.
At LanceDB, I help build the retrieval and data layer for multimodal AI. My work spans reranking, hybrid and multivector retrieval, developer tooling, and research infrastructure for video and world models. The aim is to make advanced retrieval useful as a product primitive, not a collection of disconnected techniques.
At Ultralytics, I helped build and scale the machine learning systems behind one of the largest open-source computer vision ecosystems. I led the ML team through YOLOv5 segmentation and YOLOv8, then worked across product, community, and business development to make the ecosystem easier to use and sustainable to build.
At Weights & Biases, I built the open-source integrations and partnerships channel. Making experiment tracking native to tools such as Transformers, YOLOv5, and Catalyst allowed W&B to meet developers inside the workflows they already used. The program became the company's strongest user-acquisition channel.
I use independent research to explore ideas before their product shape is clear, and open source to make the useful parts inspectable and reusable. Current threads include reproducible infrastructure for world-model research and experiments on rotary position embeddings for contrastive vision-language models. I also publish the code, working notes, and explanations that come out of the process.