I finally got tired of trusting raw LLM outputs in production. So I built the firewall to fix it. It scores hallucination risk in real-time before the user ever sees the text. No more blind RAG…