About

Helping build the retrieval and data layer that multimodal AI products can rely on.

• Making retrieval more expressive: built the reranking framework in LanceDB, hybrid search, and multivector and ColPali support so applications can combine different retrieval signals instead of committing to a single search method.
• Connecting research to working systems: published practical studies on training custom rerankers and fine-tuning embedding models, with the trained artifacts on HF Hub.
• Making world-model experiments reproducible: co-authored stable-worldmodel and built its Lance-based video data layer, streaming dataset writers, and dataloader benchmarks. The goal was to make data movement a shared, measurable part of the research.
• Reducing adoption friction: brought LanceDB into the tools where developers were already building, including LlamaIndex, LangChain.js, AutoGen, Letta, Agno, FiftyOne, and Meta's synthetic-data-kit. I also created VectorDB Recipes and YOLOExplorer as working examples rather than abstract documentation.

Led the ML team and helped build Ultralytics into the world's most popular open-source computer vision framework, now with 60,000+ GitHub stars.

• From models to a platform: co-authored the YOLOv8 model family and helped build the Ultralytics Python framework, giving detection, segmentation, classification, and pose a consistent interface across training and deployment.
• More capability on ordinary hardware: shipped YOLOv5 instance segmentation, the first instance-segmentation model to run in real time on a CPU.
• Research code people could depend on: added classification support, rebuilt the experiment-logging architecture, and hardened distributed and fault-tolerant training.
• From library to sustainable product: turned user feedback into features such as Explorer, a semantic and SQL interface for understanding computer vision datasets. I also led community and business development work that created the company's first revenue from its open-source product.

Used integrations and partnerships to make a developer product part of the open-source ML workflow.

• Integration as distribution: built and led the open-source integrations and partnerships channel, embedding W&B experiment tracking natively in Hugging Face Transformers, including its hyperparameter-sweeps backend, as well as YOLOv5, Catalyst, Tianshou, and CycleGAN/pix2pix.
• What changed: developers could use experiment tracking without leaving their existing tools or maintaining custom glue code. The program became the company's highest-performing user-acquisition channel.
• Partnership as product development: worked directly with ML tool authors to understand their users and shape the integrations.
• Explaining the underlying ideas: wrote tutorials on distributed hyperparameter optimization and pipeline versioning, including a Towards Data Science study of scalable HPO methods.

• Reproducible world-model research: contributed the video data layer to stable-worldmodel, a shared platform for comparing world-model ideas without rebuilding the surrounding system for every experiment.
• Alternative representations for vision-language models: studying rotation encoders under the SigLIP objective, including controlled ablations and training-free long-context extrapolation.
• Research impact: the papers and software releases I've co-authored have received 17,000+ academic citations.
• Research as a working record: publish code, artifacts, negative results, and literature reviews so an experiment can be understood and extended, not only cited. The current threads are on the research page.

• Improving RAG by reranking and fine-tuning: why retrieval quality often depends more on the embedding and reranking stages than on adding another generation technique.
• Hybrid search as a system: how lexical and semantic signals can complement each other, and where reranking belongs in the retrieval pipeline.
• Learning by rebuilding: more than 30 videos that take apart ML systems, computer vision models, and retrieval methods from first principles on YouTube.

For event organizers, ready to copy:

• Ayush Chaurasia builds AI products and pursues independent research on multimodal learning, retrieval, world models, and alternative architectures for intelligence. At LanceDB, he works on retrieval and data infrastructure for multimodal AI. His earlier work spans open-source computer vision, developer tools, and growing technical products through open-source ecosystems. His co-authored papers and software releases have received more than 17,000 academic citations. He speaks about turning machine learning ideas into systems people can build with.