Natural-language assistant for corporate databases

Problem

Every business question went through the one person who knew SQL and the schemas of several databases. Answers took days, and nobody could check how a number had been produced.

Self-hosting on the company's own GPUs was a requirement: customer data could not be sent to an external API.

Approach & architecture

• Researched natural-language-to-SQL approaches first and had the plan reviewed by a panel of five reviewers before implementation.
• Connectors for PostgreSQL, MySQL and SQL Server read schemas, keys and sample values into a catalogue; human-reviewed descriptions are kept apart from AI drafts.
• Every query is parsed and checked before it runs — SELECT only, allowed functions, enforced row limits — and a connection whose account can write is refused until a person accepts it.
• The LLM is reached through any OpenAI-compatible endpoint, so the same system runs on a cloud model or on local GPUs.

Result

Answers stream with the SQL, a sortable table, CSV export and charts; users mark good answers as verified to grow a trusted query library.

What made it robust

• An evaluation set with baselines and a gate command, so prompt or model changes cannot silently lower quality.
• Stored credentials encrypted with AES-256-GCM; read-only database role scripts for all three engines.
• 100+ automated tests.