Agentic Text-to-SQL in Production: What 500+ Data Marts and 30M Users Taught Us About LLM Orchestration
Hard-won lessons from building and operating an on-prem NL2SQL platform serving 30M+ users across 500+ data marts with open-weight LLMs — schema retrieval and routing across hundreds of marts, SQL generation with self-correction loops, and guardrails against hallucinated queries. Target audience: data engineers, platform engineers, and anyone moving an LLM agent from demo to dependable enterprise infrastructure.