Project
DuckPilot
Natural-language SQL over your CSVs

Overview
A browser-based analytics desk that lets you drop a CSV or Parquet file and ask questions in plain English. Non-technical users lose hours wrestling with spreadsheets and ad-hoc SQL just to answer basic data questions; DuckPilot turns that into a sentence.
Architecture & system design
React and Next.js frontend with DuckDB compiled to WebAssembly, so every query runs client-side and the rows never leave the browser. Questions route to OpenAI, Claude or Gemini with SQL fallback strategies when a query fails, and WebWorkers keep the UI responsive on large files. No backend or infrastructure at all.
Key deliverables
- Drop a CSV or Parquet file and ask questions in plain English
- DuckDB-WASM runs every query client-side; rows never leave the browser
- Routes to OpenAI, Claude or Gemini, with SQL fallback strategies when a query fails
- WebWorker support keeps large files from blocking the UI
- Generated SQL and results are shown and validated against the source data
- Multi-hour manual analysis becomes minutes, with no signup, no API key and no backend
Engineering decisions
Built with a team. The decision that shaped the product was compiling DuckDB to WebAssembly: the analysis stays entirely in the browser, which removes the backend and the privacy trade-off in one move. The generated SQL is surfaced and checked against the source data, rather than asking people to trust an answer they cannot see the working for.