TriageAI Coach

Inspiration

250-440k annual deaths from preventable medical errors. 12 million U.S. adults experience a diagnostic error each year. Healthcare professionals rarely receive structured feedback on their clinical reasoning, with traditional training relying on sporadic attending physician feedback during high-pressure rotations. We recognized that Claude's sophisticated reasoning capabilities could simulate an expert attending physician for triage training, making high-quality clinical reasoning coaching accessible 24/7.

What it does

Our AI Attending Physician evaluates triage case presentations against 9 expert clinical competencies, then generates targeted Socratic questions that guide learners to discover gaps in their own diagnostic reasoning and patient prioritization skills. The system provides adaptive, real-time feedback creating a safe environment for practicing triage presentations with expert-level feedback grounded in 200,000 PubMedQA data points retrieved through our RAG system.

How we built it

We started with an agent builder prototype then converted to the Claude API to design a four-agent orchestration system where specialized agents evaluate presentations, generate Socratic questions, assess understanding, and manage adaptive conversation flow. We used PubMedBERT to encode 200,000 PubMedQA data points into a PostgreSQL database, then created a RAG system to retrieve relevant medical documents that ground our model's feedback in evidence-based literature. The platform features a React + Vite frontend deployable to GitHub Pages and a FastAPI backend with PostgreSQL persistence, designed for scalable deployment on managed PaaS platforms like Render or Railway.

Challenges we ran into

Designing the multi-agent architecture required careful orchestration of state flow across four specialized agents, ensuring each agent had access to the right context without overwhelming token limits. Embedding 200,000 PubMedQA documents with PubMedBERT was computationally intensive and time-consuming, requiring careful batch processing and error handling to complete the full dataset. The DevOps pipeline was complex, requiring us to coordinate the GitHub Pages frontend deployment, set up managed PostgreSQL databases, and configure the FastAPI backend on cloud platforms while ensuring all components could communicate securely across different domains.

Accomplishments that we're proud of

We built a production-ready multi-agent system with sophisticated orchestration where each agent has clear responsibilities and manages complex state transitions seamlessly. Our 9-metric evaluation framework is grounded in medical education literature, paired with a fully functional RAG pipeline that retrieves evidence-based medical literature in real-time, creating genuinely personalized learning pathways with transformative potential to improve patient safety at scale.

What we learned

Breaking complex educational tasks into specialized agents made the system more maintainable, with each component excelling through iterative prompt engineering tailored to the unique demands of triage decision-making.

What's next for AI-Powered Nurse Training

Our PostgreSQL-based architecture is designed to scale and support thousands of simultaneous users, with response caching and optimized vector search to keep costs manageable while maintaining high-quality feedback. We envision potential partnerships with nursing schools and hospitals as a subscription service, with the goal of becoming a standard triage training tool across healthcare institutions by offering accredited continuing education credits and expanding our RAG database with specialty-specific medical literature.