Self-hosted multi-agent AI platform: generates, evaluates and refines LLM answers until they hit a quality bar (~22% more accurate, ~45% fewer hallucinations)