a live decision engine where a council of specialized LLM agents (analyst, contrarian, sentinel) debates high-stakes questions, votes, and explains every verdict. Every juror is served locally by vLLM on AMD Instinct GPUs via ROCm, with a real-time dashboard showing per-juror tokens/sec, latency, and GPU utilization — "multi-agent AI you can audit," plus a benchmark story that shows what AMD's compute stack can actually unlock