This project, AgenticLakehouse, explores the cutting edge of "Agentic Data Analytics." I didn't just want a chatbot; I wanted a "living" interface for the Lakehouse. The result is a Multi-Agent System that intelligently orchestrates tasks, from querying Unity Catalog to browsing the web, deployed directly as a Databricks App.
To achieve this, I separated the system into a "Brain" and "Hands":
• The Brain (LangGraph): A router architecture that assesses user intent and dispatches tasks to specialist agents (e.g., a Databricks Agent, a Web Search Agent, etc,.)
• The Hands (MCP): A custom Model Context Protocol server that standardizes how these agents discover schemas, inspect lineage, and safely execute Spark SQL queries
Tech Stack Breakdown:
• Orchestration: LangGraph (Router + Specialist Agents)
• Protocol: Model Context Protocol (FastMCP)
• LLM Models: Groq
• UI/App: Gradio (Databricks Apps)
• Compute: Databricks Serverless SQL, Databricks App, Render
• Observability: LangSmith
In the Medium post, I walk through the full architecture, including how to optimize Unity Catalog metadata for LLM context windows and managing state across multiple agents.
