Counsel-in-a-Box: Legal AI on MI300X

Under-resourced legal teams spend 4-12 hours per matter reading hundreds of pages of contracts to produce a single review memo. Existing AI tools chunk documents through RAG, meaning the model never sees the whole picture — a clause on page 12 that contradicts a definition on page 187 gets missed. Counsel-in-a-Box solves this by loading the entire deal package into one context window on an AMD Instinct MI300X. Five specialist agents — Clause Extractor, Risk Scorer, Obligation Tracker, Cross-Reference Auditor, and Memo Drafter — read the same full corpus and produce a coordinated, cite-checked legal memorandum. The system found real issues in SEC EDGAR M&A filings: a $22M termination fee discrepancy between sections 60 pages apart, conflicting definitions across documents, and missing exhibits — findings that RAG-based approaches structurally cannot detect. Key results: 89 tokens/sec on Qwen3-32B via vLLM on ROCm 7, 62% GPU utilization (121 GB of 192 GB), and $1.99/hr on AMD Developer Cloud vs $8.20/hr for 2x H100. The MI300X's 192 GB HBM3 enables full-context multi-agent inference without quantization, chunking, or multi-GPU overhead — an architecture that is impossible on 80 GB GPUs.