Build a Self-Hosted AI PR Reviewer | Cooking with MLOps Recipe 3

Description

Pre-packaged AI tools are quick. But they mean someone else controls the ingredients. In Recipe 3 of the Cooking with MLOps cookbook, Matt walks through how to build a self-hosted AI agent that reviews GitHub Pull Requests: inspecting each change, judging fit, and offering clear, useful notes. Consistently. On your own infrastructure, with your own model, and without sending data to third parties. The stack: - PydanticAI to build the agent (the mixing bowl) - vLLM to serve the model (the hob) — Qwen by default, but swap in any open-weight model - MLflow to version your prompts (the tasting notebook) - Pydantic Evals to test the agent's output (the tasting spoon) - Pydantic Logfire to observe it in production (the thermometer) This is Part 1: building the prototype. Part 2 covers production governance. Find the full cookbook at: https://bit.ly/4bT6DiI