Cheap Models First: Building an LLM Cascade Router That Cuts Your AI Bill Without Hurting Quality