Learn how to build and continuously optimize a defensive coordinator chatbot guided by SME feedback with Databricks Agent Framework, Unity Catalog tools, and MLflow. Surface opponent tendencies by down and distance, formation, personnel, and situation—grounded in governed data and production-quality agent operations.

databricks

Building self-optimizing AI agents requires domain expert feedback, not generic LLM judges. This implementation demonstrates a two-phase architecture: first, build a tool-calling agent using Unity Catalog SQL functions for deterministic data retrieval and LLMs for semantic interpretation; second, continuously optimize using MLflow's align() to calibrate judges against subject matter expert preferences, then optimize_prompts() with GEPA to evolve system prompts automatically. The football defensive coordinator chatbot example shows how expert labeling sessions create ground truth that shapes judge alignment, which then guides mathematical prompt optimization—translating domain expertise into concrete system improvements without developer guesswork.

Self-Optimizing Football Chatbot Guided by Domain Experts on Databricks