My LLM Judge Went From 70% to 96% Agreement With Humans. Here’s What I Learned.

An engineering diary about evaluating an AI waiter, debugging an LLM judge, and discovering that a better rubric can matter more than a cleverer model. When I started building the evaluation pipeline