Build a framework for modular medical diagnostics that strictly separates LLM conversation from Bayesian reasoning engines. Allow developers to plug in different probabilistic belief graphs.