# Show HN: InverSQL: Build SQL interactively in inverse

> Source: <https://github.com/rentruewang/inversql>
> Published: 2026-09-08 16:44:14+00:00

1. User doesn't want to write SQL.
2. User uploads CSV to [inversql streamlit app](https://inversql.streamlit.app)
3. User selects cells (that will be selected by the SQL).
4. We **overfit** a`scikit-learn` binary decision tree on the data.
5. We decompose the tree, convert to boolean logic (explainable AI part).
6. We simplify the logic with `sympy` .
7. Generate SQL from previous steps (joins to JOIN and boolean to WHERE).
8. User sees the SQL.
9. User is happy.

| 🎬 Demo in a GIF | 🏛️ Architecture diagram | 
|---|---|
|                 [Link to](/rentruewang/inversql/blob/main/assets/quick-demo.gif)                    [live demo site](https://inversql.streamlit.app)         here. |   | 

For each individual SQL query candidate (the shortest one is displayed in the UI), we need to retrain a new decision tree.

But... The decision tree fitting is honestly fast, don't worry about this.

That's pretty much it!

If you have read this far, please consider giving me a [star (⭐)](https://github.com/rentruewang/inversql/stargazers) or a fork (🍴).

This will keep my motivation going!

Or if you have too much cash at hand: 

If you **REALLY** like my work, nowadays I'm working on [`aioway`](https://github.com/rentruewang/aioway),
it's an automated training and inference engine that does the following:

- Adapt to hardware it runs on (optimal hardware usage)
- Adapt to data it trains on (figure out architecture on its own)
- Incremental training (never overfit or underfit)

Contribution welcome!

To contribute, refer to [CONTRIBUTING.md](/rentruewang/inversql/blob/main/CONTRIBUTING.md),
and our [CODE_OF_CONDUCT.md](/rentruewang/inversql/blob/main/CODE_OF_CONDUCT.md).

Inspired by [regexgen](https://github.com/devongovett/regexgen#how-does-it-work)'s process. Instead of regex we do SQL. Instead of selecting text we do select records. Decision tree is my inspiration tho.