Chang She is the co-founder and chief executive of LanceDB and an early co-author of pandas. After helping make Python foundational to modern data analysis, he now builds infrastructure for the images, videos, embeddings, and training datasets that conventional tabular systems struggle to handle.
She began his career in quantitative finance before joining pandas creator Wes McKinney to work on the library full time in 2012. He also developed quantitative portfolio-management software and co-founded DataPad, a visual-analytics startup where he served as chief technology officer. The DataPad team joined Cloudera in 2014, extending his work into enterprise-scale data infrastructure.
At Tubi, She became vice president of engineering, working on personalized recommendations, machine-learning infrastructure, and experimentation. Those systems exposed a growing mismatch between established tools for structured data and the multimodal datasets required by production AI.
With Lei Xu, he founded LanceDB and developed Lance, an open-source columnar format built for AI workloads, alongside the LanceDB database. In June 2025, he announced a $30 million Series A led by Theory Ventures and an expanded multimodal lakehouse strategy. The company reported more than 20 million open-source package downloads and identified Runway, Midjourney, and Character.ai among its enterprise customers.
- One format for three demanding workloads. AI training requires fast scans to select examples, fast random access to shuffle or retrieve individual records, and efficient streaming of large images or videos. She designed Lance to support all three together, reducing the need to duplicate datasets across specialized systems. His random-access benchmarks connect that design directly to search, feature retrieval, and training.
- A lakehouse that supports both training and retrieval. She combines object storage with separated compute and storage so the same underlying dataset can support SQL analytics, vector and full-text search, preprocessing, model training, and low-latency production queries. His work on enterprise retrieval infrastructure emphasizes multimodal storage and serving architecture; Harvey-specific legal evaluation and confidentiality practices belong to collaborator Calvin Qi.
- Datasets that change without wholesale rewrites. Lance supports zero-copy schema evolution, automatic versioning, and rollback, allowing teams to add embeddings, labels, or experimental features without copying entire datasets. Efficient row-level access also enables training pipelines to shuffle references instead of moving large underlying records.
- Interoperability as practical infrastructure. Lance integrates with Apache Arrow, Spark, Ray, pandas, Polars, and PyTorch. She's account of adding Windows support also credits scientific-Python maintainers and the engineers responsible for the implementation, reflecting his longstanding attention to packaging, compatibility, and usable open-source foundations.
