Speaker : Mike Taylor author of book "Prompt Engineering for Generative AI: Future-Proof Inputs for Reliable AI Outputs" Event: London Agentic AI Meetup (25th September 2025) Talk: The Evaluator–Optimizer Pattern in DSPy with GEPA Link to the Slides and Code in the Comments section below. In this talk Mike demonstrated how to train an LLM-as-a-Judge and then use it to optimize fuzzy generative tasks. This demo teaches the evaluator-optimizer pattern in DSPy by building a joke-telling AI system that demonstrates how to create, evaluate, and optimize AI programs systematically. The example shows how to build a comedian AI that generates jokes in specific comedian styles and uses an AI judge to evaluate and improve joke quality through evolutionary GEPA optimization. - Core DSPy Concepts Demonstrated 1. Signatures and Predictors: Creating custom input/output specifications for AI programs 2. GEPA Optimizer: Using evolutionary algorithms for prompt optimization : Implementing AI-based evaluation metrics 4. Evaluation Framework: Building systematic performance measurement systems - Key Technical Components 1 Comedian Program: Generates jokes in specific comedian styles (Ricky Gervais, Billy Connolly, etc.) 2. AI Judge: Evaluates joke quality using LLM as Judge 3. GEPA Evolutionary Optimizer: Automatically improves both comedian and judge performance 4. Dataset Creation: Building training datasets with funny vs unfunny jokes Performance Evaluation: Measuring and comparing system improvements