Blackwell - Deep Agent Clinical Decision Support System
Lukas Torquato
Published
Developed a sophisticated multi-agent AI system designed to streamline the clinical intake process and provide practitioners with data-driven diagnostic hypotheses and treatment plans.
Architecture: Engineered a coordinated multi-agent pipeline using LangChain and LangGraph, featuring specialized agents for patient anamnesis, clinical evaluation, and therapeutic research.
Knowledge Integration: Implemented Retrieval-Augmented Generation (RAG) using ChromaDB to index the Medline database, coupled with a real-time PubMed web-crawling agent to retrieve state-of-the-art treatment guidelines.
Performance & Validation: Rigorously evaluated the system against a dataset of 250 simulated patient cases, achieving a 90.1% total diagnosis accuracy and a mean clinical quality score of 83.4/100.
Safety & Compliance: Designed an “AI Patient” simulation framework for testing and integrated emergency detection protocols to identify critical symptoms (e.g., chest pain, slurred speech) during the anamnesis phase.
Tech Stack: Python, Google Gemini, Ollama (Local LLMs), FastAPI, ChromaDB, LangGraph