Vol. 01 — Multi-agent data analyst
DataPilot is a team of agents that plans your question, writes three SQL candidates, lets them vote, runs the winner read-only, and checks every number in the answer against the data. Open any answer and see exactly what each agent did.
FIG. 1 — THE PILOT'S INSTRUMENT
Try one
A ranking with year-over-year change, a statistic computed by Python in your browser, and a vague question the planner refuses to guess at.
Have your own spreadsheet? Upload a CSV, Excel, JSON or SQLite file and ask it anything — private to you, deleted automatically.
The crew
Not one prompt that guesses SQL. A supervised LangGraph team with parallel candidates, self-repair, human checkpoints, budgets and stop rules — every step traced.
See the architectureBreaks the question into steps — or asks you what you meant.
Finds the right tables, columns, join paths and the exact spelling of values.
Three strategies write candidates in parallel: direct, plan-then-SQL, few-shot.
Parses every statement; anything but one bounded SELECT is rejected.
Read-only connection, read-only files, a 5-second timeout and a row cap.
Candidates vote by result; a checker confirms columns, filters and grain.
Writes pandas that runs in your browser — Python never runs on the server.
Emits a validated Vega-Lite spec; the data is injected by your browser.
Writes the answer in under 120 words, from the results only.
Checks every number against the data and removes anything unsupported.
A SQL parser, a read-only engine connection and read-only files. Even a guard bug can't write.
Analysis code runs in Pyodide (WebAssembly) in a Web Worker with network access removed.
The answer's numbers must come from the results — or a derivation code can recompute. Else they're cut.
Measured, not claimed
A fixed, stratified 150-question subset of the public BIRD Mini-Dev benchmark, an ablation that turns each agent on one at a time, and a cost-vs-accuracy chart — with 95% bootstrap intervals and the failures left in.
Open the benchmarks48.0%
Single-call baseline · n=50
48.0%
C0 best config · n=50