How to build evaluation systems for LLM-powered features — covering human eval, automated checks, LLM-as-judge, eval datasets, and regression prevention.