Problem
CAVA, a restaurant chain, centralized its data in one Databricks platform (about 3,000 tables) and invited business users in. They immediately asked what the data meant and where to find it. The small data team became the bottleneck, and decisions were made without answers. Genie rooms (Databricks' domain Q&A spaces) helped, but one room holds few tables and has a limited context window, so it could not cover the estate.
Example from the article: a finance user wants Q1 revenue by region against budget and has to message an analyst and wait a day.
Solution in brief
CAVA built Ask Astro, a supervisor agent behind a chat-like Databricks app. A routing LLM sends each question to a domain Genie room or an MCP tool and returns one answer under the user's own permissions. An LLM judge scores production traces daily, with an alert below 85% accuracy.
Questions this case answers
- How to route business questions across several scoped text-to-SQL spaces and tools with a supervisor agent that enforces each user's own data permissions
- How to monitor a production data agent every day with an LLM judge and alert when accuracy drifts below a threshold
- How to let business users ask data questions in plain English without waiting for an analyst
- How to route questions across several Genie rooms and tools with a supervisor agent
Tools and methods
- Agent Bricks supervisor — routing LLM that picks the tool or room and synthesizes the answer
- Genie rooms — one per domain; turn questions into SQL over governed tables
- MCP tools (Atlassian, Tavily) — Jira and Confluence access, web search
- Lakebase (Postgres) — conversation history and user feedback
- MLflow tracing — replayable trace of every supervisor call
- LLM-as-a-judge — custom scorers run daily on production traces
- AI/BI dashboard — cost, latency, activity and satisfaction
Google or a one-time email code. No password.
