Evaluating and Optimizing Educational Content with Large Language Model Judgments