Unstructured clinical text poses major challenges for automation and interpretability in healthcare. Hermes is a multi-agent large language model (LLM) framework that sequentially converts discharge summaries into structured SOAP-format reports and knowledge graphs.
The system integrates four specialized agents — Hermes-R (report generation), Hermes-G (graph construction), Hermes-Q (question-answer generation), and Hermes-A (answer validation) — that collaborate through iterative refinement and semantic validation loops to enhance factual accuracy and consistency.
Accepted and presented at the AAAI SecureAI4Health 2025 Symposium (poster + lightning talk), Hermes demonstrates how coordinated LLM agents can improve the reliability and interpretability of clinical natural language processing.