NL-to-SQL copilots are easy to demo and hard to productionize: they hallucinate joins, leak restricted fields, and confidently return “correct-looking” numbers. This talk gives you a practical blueprint to ship an analytics copilot that your data team can actually trust—by combining policy enforcement (guardrails) with evaluation harnesses (evals) and failure-aware UX.

You’ll learn a concrete architecture for (1) permission-aware retrieval of semantic context (metrics, dimensions, join rules), (2) constrained generation that produces explainable SQL and bounded outputs, and (3) automated evals that catch regressions before your users do. You’ll leave with checklists, prompt and policy patterns, and an “MVP eval pack” you can run locally—no live internet required.

What you’ll walk away with

A reference architecture for NL-to-SQL copilots (semantic layer → policy checks → SQL generation → verification → answer)

A minimal eval suite (accuracy, safety, “metric integrity,” and regression gates)

Guardrail patterns for blocked columns, row filters, PII, and prompt injection

A playbook for “safe failure”: when to refuse, ask clarifying questions, or route to a human