Quantile

A chat-first analytics workbench: connect a SaaS account, ask a question, and get a four-stage pipeline you can read, edit, and re-run.

Overview

Quantile turns a plain-English question about your app data into a working analytics pipeline. Connect a SaaS account (Airtable first), ask something like "extract monthly revenue and build a dashboard", and the system writes and runs four stages: extract, load, transform, visualize.

Every stage is a real file in the workspace. Nothing is hidden behind a chat bubble.

The problem

• Answering a business question from SaaS data usually needs a chain of people and tools: someone to export, someone to model, someone to chart.
• Classic BI tools assume the data is already modeled.
• Chat-with-your-data tools hide their queries, so answers can't be checked or reused.

How it works

• Extract: a generated Python script pulls records through the connected account and writes Parquet.
• Load: data lands in DuckDB, in the browser via DuckDB-Wasm where possible.
• Transform: SQL files shape the data for the question at hand.
• Visualize: a dashboard spec renders with Mosaic vgplot.
• Follow-ups are routed to the narrowest layer that can answer them. "Break revenue down by category" edits the transform, not the extractor.

Design choices

• Every stage is an editable file, so the user can always see and change what ran.
• Row data stays out of the model prompt and mostly off the server. The LLM sees schemas, not customer records.
• Untrusted generated code executes in Vercel Sandbox, never on the app server.

Stack

• FRAMEWORK: Next.js, React, TypeScript
• EXTRACTION: Python (uv), Pipedream Connect
• ANALYSIS: DuckDB + DuckDB-Wasm
• DASHBOARDS: Mosaic vgplot
• AI: Vercel AI Gateway
• RUNTIME: Vercel Sandbox, Vercel Blob
• DATA: Postgres + Drizzle, Neon previews, NextAuth