
Boardroom is a multi-agent document intelligence system that goes beyond summarisation. Instead of a single LLM response, it runs a structured three-phase debate and verification workflow over any uploaded PDF or TXT document. Phase 1 runs three agents in parallel on AMD Instinct MI300X — the Analyst extracts key facts with page citations, the Skeptic aggressively challenges every claim, and the Strategist produces actionable recommendations grounded in evidence. Phase 2 is a live debate — the Skeptic reviews the Strategist's output and challenges it again. The Strategist fires back with a revised position. Phase 3 is the final audit — the Auditor reads the entire debate, marks any unverified claims with [UNVERIFIED], and delivers a single consolidated verdict. The system runs Llama 3.1 70B FP8 on a single AMD Instinct MI300X with 192GB HBM3 memory via vLLM with prefix caching enabled. This allows all four agents to share the same document KV cache, making parallel execution genuinely faster rather than just concurrent. The full document is read without chunking — no RAG, no retrieval, just direct megacontext reasoning. Built with Python, Gradio, and the OpenAI-compatible vLLM endpoint on AMD Developer Cloud.
10 May 2026