Built a privacy-first approach to AI-assisted data analysis. The challenge: how do you let AI help with SQL queries without exposing sensitive data?
Solution: AI only receives table schemas (column names + types). It generates queries based on structure alone, which run locally via DuckDB WASM.
All processing happens client-side. Supports CSV, Parquet, JSON, Excel up to several GBs or bring from HuggingFace, S3, etc and start asking questions.
Try here:
https://datakit.page
Curious if others have tackled similar privacy constraints with LLMs.