NL-to-SQL in Six Weeks

A fintech wanted business users to query their data in plain English. They'd tried off-the-shelf tools and custom scripts but hadn't found something accurate, fast, and safe enough for production. We ran a short discovery, then built an MVP in six weeks.

What we built

• A clear separation between the natural-language-to-SQL layer and the execution layer.
• Strict read-only access for generated queries.
• A human-in-the-loop step for any query touching sensitive tables.
• Documentation, tests, and a simple UI so internal teams could use it in a sandbox.

Why it worked

We didn't spend months designing the perfect architecture. We made sensible choices for version one and shipped. User feedback from the pilot drove the next sprint: we improved handling of ambiguous questions and added better explanations for the generated SQL. The product is now in use across multiple business units. For organisations that want to move fast without sacrificing rigour, this combination of aligned strategy, iterative delivery, and strong engineering practice remains the most reliable path to impact.