20 November 2026
Building in Public/Building an AI data analyst: the series
A series documenting how an AI data analyst is built from scratch with coding agents: architecture, data connections, trust, testing and release.
Most AI demos can answer a question. Very few answer it in a way a finance team would sign off on.
That gap is why I spend my own time helping build an AI data analyst: you ask about business data in plain language and get an answer you can check. You see the SQL it ran, the numbers it found and the chart it drew.
Most of the engineering is done together with AI coding agents. So this is also a real experiment in building serious software with agents: what gets faster, what breaks, and what still needs a human.
The series
- Why a trustworthy data assistant? The problem, and what already exists.
- Architecture. Interface, backend, connectors and the model layer.
- Connecting to real data safely. Snowflake, BigQuery and Databricks with read-only access.
- From a question to correct SQL. Why “looks right” is not enough.
- Using the models people already have. See Using Claude, Codex and Gemini through their command-line logins.
- Making agents check their own numbers. Catching wrong answers before users see them.
- From chat to report. Charts, files and exports.
- Testing with real data. See Testing AI features with real data, not mocks.
- Building most of it with coding agents. What worked and what did not.
- Shipping a desktop app. Signing and notarising on the Mac.
- Open-sourcing it. License, secrets in git history, and launch day.
The rule for this series
Real work only. Every post shows something that exists, with the trade-offs actually made, including the mistakes. No predictions about where AI is going; plenty of people write those already.
If you build with AI and want the practical version, follow along.