QueryLab
A local workspace for generating datasets, exploring them with SQL, and evaluating queries against reviewed references.
On this page
Overview
QueryLab is a local workspace for generating datasets and testing SQL. It started as an interview practice tool and expanded to support open-ended exploration, query comparison, and repeatable evaluation.
Start with a dataset idea, a business question, or a company-inspired scenario. Generate synthetic data, inspect its structure, and run queries in DuckDB. Practice questions are included by default, with a data-only option for free exploration.
How It Works
- Describe the data or question, then review the scope before generation.
- Explore the tables, run SQL, and save queries alongside the dataset.
- Compare query variants against the same saved data.
- Evaluate results using a reference query you have reviewed.
- Reopen saved datasets and evaluation reports without regenerating them.
What I Learned
Started with SQL interview practice to learn how to work with different databases and data types. Wanted to explore the data before answering a question, like in an actual work environment. That led to saved datasets and a workspace where I could write my own queries and compare different approaches.
Added evaluation features along the way to help compare and test queries. Looking to expand the agent and data generation layers further next.
Key Capabilities
- Synthetic datasets: Generate data around a business scenario using Codex or Claude CLI.
- Saved workspaces: Keep dataset snapshots, questions, and named SQL queries together.
- Query comparison: Inspect differences in columns, rows, duplicates, NULLs, and ordering.
- Reviewed evaluation: Test queries across selected datasets and compare saved runs.
Project Status
The local implementation supports dataset generation, exploration, comparison, and evaluation. File uploads, direct AI API connections, and a hosted demo are planned.
Company-inspired datasets are synthetic and unofficial. Evaluation checks results against reviewed references on selected datasets; it does not prove correctness for every possible input.