An LLM that reads your data — with no server.

A real language model runs entirely in this tab via WebGPU (WebLLM). Paste a small dataset, ask a question or request a data-quality read, and the answer is generated on your device — no API key, nothing leaves your browser. The model is downloaded once (cached after), so the only cost is your own GPU.

Your data (CSV or a list)

Checking WebGPU…

Example output · load the model to run it liveThis looks like a stops table with columns: atco, name, locality, status, modified. Likely data-quality issues: row 3 has empty coordinates (0,0); rows 4 and 7 are status "inactive" but still present; "modified" dates in rows 2 and 6 are 8+ years old (stale); "Bank of Engalnd" looks like a misspelling of "Bank of England". I'd flag inactive-but-present and stale records first.

How it works. WebLLM compiles a quantised small model (Qwen2.5-0.5B) to WebGPU and runs inference in a web worker — the same shift that's making the browser a real AI runtime in 2026. The dataset is passed in the prompt; the reply streams back token-by-token, locally. Great where data can't leave the device (regulated, on-prem, offline). Built by John Mikel Regida.