Data Expert: A Local-First Workbench for Versioned, Privacy-Preserving Dataset Exploration and AI-Assisted Transformation

Data Expert is an open-source, local-first workbench that lets one person browse, search, profile, edit, version and transform datasets from kilobytes to many gigabytes in a web browser, without uploading data anywhere. This paper describes three design contributions. First, a lazy, pending-edit architecture opens a 3GB, 1-million-row Parquet file in 0.07 s by reading only metadata, and commits changes through atomic file replacement. Second, a git-like version history for data stores row edits as reverse deltas of a few kilobytes regardless of file size, and stores transforms as hardlinked snapshots that cost no extra disk space at creation. Third, an exact full-text search over very large files uses a compact block index of per-block n-gram fingerprints plus a word list, staying under 1/100 of the file size (25 MB for a 3 GB file) while returning exactly the results of a full scan. A fourth element, an AI assistant, turns plain-language requests into plans over a library of typed functions. It sends the model only the schema, summary statistics and truncated samples, and runs every plan, and any model-written function checked by an AST sandbox, locally. We report author-measured timings on a single 8-core desktop and describe the automated test suite and the system's limits, including that its sandbox is not a hardened security boundary.