Work

• I built the retrieval features in LanceDB: reranking, hybrid search, and multivector retrieval. That's the stuff people come looking for once plain vector search stops returning what they wanted.
• I did the ecosystem integrations, because nobody is going to rebuild their stack just to try your database: native Lance support on the Hugging Face Hub, LlamaIndex, LangChain, AutoGen, and the rest of the RAG and agent stack. I also built a pile of public demos searching everything from all of Wikipedia to a billion-document web corpus, and Semantic.art, which finds artworks by mood and description rather than by title or artist.
• I wrote a lot of the company's technical content, including LLMs, RAG, and the missing storage layer for AI and the practical guides on training rerankers and fine-tuning embedding models. I usually end up writing the post I wanted to find when I was working the thing out myself.
• I kept doing peer-reviewed research along the way, including stable-worldmodel (NeurIPS 2026) with Yann LeCun and others. The rest is on my Scholar page.

• I co-authored the model families the framework is known for: YOLOv8, downloaded more than 20 million times in its first year, and the first instance-segmentation model to run in real time on a CPU.
• I started the Ultralytics framework, co-authored it and led its development, and it ended up being the world's most popular open-source computer vision framework, with 60,000+ GitHub stars, and most of my 20,000+ academic citations come from the papers and software releases I co-authored there.
• I started the community on Discord, and worked on the business development that produced the company's first revenue from open source.

• I created the open-source integrations and partnerships channel, which in practice meant getting W&B built into the tools the ML world was already using: Hugging Face Transformers, spaCy, YOLOv5, Catalyst, ESPnet, and more. It ended up being the company's biggest source of new users.
• I wrote a lot of research engineering pieces. Some of them are sadly gone now, but the ones still up are in the blog section.
• I also worked on the developer experience of the product itself, mostly by meeting users where they already were: smarter login timeouts on platforms with no interactive terminal, and the right defaults for specific verticals like computer vision.

For event organisers, ready to copy:

Ayush Chaurasia builds AI products for developers and researchers and does independent research in multimodal learning and world models. He is the founding ML engineer at LanceDB, previously led the ML team at Ultralytics, where he started and led the Ultralytics framework, and built and led the open-source integrations and partnerships channel at Weights & Biases. His co-authored papers and software releases have received more than 20,000 academic citations. He speaks about retrieval, multimodal systems, and building developer products.

If you want to get in touch about a talk, or anything else, I'm at [email protected].