I reject the premise that you can’t use LLM’s to build reliable features.