Supervisor models that review AI answers before users see them

A single AI model can miss things or make errors, and output quality varies when nothing reviews it. I built a multi-model system where worker models answer and supervisor models review and validate the answers before they reach a user, to catch hallucinations and inconsistencies. It runs Claude, GPT and other providers side by side, streams answers live and lets you set each agent's role and limits. It was built so AI output is constrained, observable and repeatable.