AI Tumor Board

Tumor boards are where cancer treatment decisions are made — radiology, pathology, medical oncology, and surgery in one room. They are also a coordination bottleneck: scheduling 4+ specialists takes weeks, clinical reasoning is discussed verbally but rarely captured, minority opinions get lost, and there is no auditable trail for governance. AI Tumor Board models that meeting as six AI agents collaborating in a single Band room. A script-driven coordinator invites each specialist one at a time - Radiologist, Pathologist, Medical Oncologist, and Surgeon, each providing a structured opinion with staging, biomarkers, treatment recommendations, and confidence levels. After all specialists respond, an independent Safety Officer reviews the full case and can clear, request more information, or issue a hard veto to block autonomous consensus. The system uses three different Band adapters (PydanticAI, Claude SDK, Codex), three providers (AI/ML API, Anthropic, OpenAI), and multiple models (gpt-4.1, claude-sonnet-4-6, qwen3-32b). The Safety Officer deliberately runs on a different framework and provider than the specialists it reviews, decorrelating blind spots by design. Key design decisions: dissent is a mandatory output field, a board that hides minority views is worse than no board. Every recommendation carries requires_human_review=true and human_signoff=null. The Band room transcript is the system of record — no separate database. The export command reconstructs the full board note directly from the transcript. The project includes a Streamlit UI for one-click demos, four synthetic cancer cases (NSCLC, triple-negative breast, colorectal with liver mets, borderline resectable pancreatic), and a nudge-retry mechanism that handles models that provide prose but skip the structured format. All cases are synthetic. This is decision-support software, not for clinical use without proper governance and regulatory review.