The client built a secure code-execution sandbox product and needed a compelling, real-world demo: database analytics for non-technical users. Non-technical users normally wait on someone else to write and run SQL — often a multi-day turnaround. Letting an LLM write and run the query itself introduces a second problem: running AI-generated code directly against production infrastructure is a real security risk.

A Flask-based AI platform that lets a user ask a database question in plain language and get back an answer — with the AI-generated code never touching anything but a secure, isolated sandbox.

Natural language → SQL/Python — user questions are converted into executable database queries or Python data-processing code. MCP (Model Context Protocol) structures how the AI layer interacts with the underlying data and tools.

Sandboxed execution environment — the piece that actually solves the security problem: AI-generated code runs inside an isolated sandbox, separate from the client's real infrastructure, instead of executing directly against production. Automated visualizations turn raw query results into something immediately readable.

The interesting engineering decision wasn't the natural-language-to-SQL part — that's an increasingly common pattern. It was recognizing that "let the AI run code" and "let the AI run code safely" are two completely different engineering problems, and solving the second is what made the demo credible to a technical audience.