TriageFlow | Devpost

Inspiration

One of our teammates went to the ER with a broken arm and waited almost 6 hours before. In busy experiences like his, even the smallest of inefficiencies can add up. We were inspired by the idea that digital tools and automation can support nursing staff, not replace them, by monitoring evolving conditions, surfacing early warnings, and giving transparent, actionable recommendations. Our project builds on this by combining patient data (history, current symptoms, images) with a speaking agent, queue-visualisation and prioritisation logic that adapts as conditions change.

What it does

Triage score determination: calculates a priority/level based on current symptoms, history, images of injuries, patient comments. (Follows ESI guidelines)
Patient records retrieval (HL7 FHIR): retrieves prior visits, known conditions and reason for visit so the nurse and system have full context.
Current symptoms & patient comments capture (speaking-agent interface): patients can speak or type their current symptoms and comments and the system prompts for clarifying details.
Nurse dashboard: presents the queue, allows manual add/remove of patients, queue re-ordering, and re-running of triage logic.
Notification system: if the triage suggestion increases (i.e., higher risk), the system triggers an alert to the nurse (and optionally the patient) with a new recommendation.
Patient view: Specific links for each patient sent to their phone number, allowing patient to update their symptoms in the system through a chatbot interface.

How We Built It

Fetch.ai UAgents — created a smart agent that orchestrates data flow between the patient intake form, database, and AI model.
ChromaDB — stores synthetic patient medical data (HL7 FHIR), enabling efficient semantic search and retrieval of relevant patient histories.
Twilio — used for SMS communication to text patients links so they can access their place in the triage queue and receive updates.
Claude API (Anthropic) — processes patient records and symptoms to produce triage summaries and risk-level recommendations.
Frontend: JavaScript + React
Backend: Python + Flask

Challenges We Ran Into

Integrating LiveKit for NLP on the frontend — managing rooms, sessions, and audio streaming was complex.
Developing a custom priority queue — ensuring that triage scoring and patient order dynamically updated in real time.
Prompt engineering for consistent feedback — tuning prompts so the AI agent produced stable, actionable outputs each time.
Working with AI agents for the first time — learning how to manage message flow, context retention, and error recovery.

Accomplishments That We're Proud Of

Developed a fully custom priority queue system that supports real-time triage reordering.
Achieved consistent AI-generated feedback using advanced prompt-engineering and model tuning.
Successfully integrated all components — backend, agent, triage logic, and frontend — into a cohesive working prototype.

What We Learned

Prompt engineering with Claude 3.0 — designing structured prompts to yield consistent, context-aware triage recommendations.
Organizing and querying patient data in ChromaDB — leveraging vector embeddings for efficient, semantically relevant searches.
Breaking projects into subtasks — improving collaboration, productivity, and version control across multiple team members.

What's next for TriageFlow

Deploy backend and agent services on a secure, scalable cloud infrastructure.
Integrate live hospital databases to access authentic patient records and enable real-time updates.
Fine-tune waiting time predictions using queue length, triage severity, and patient flow analytics.
Refine chatbot prompts to deliver more empathetic, context-aware responses for distressed patients.
Expand multilingual and accessibility support to better serve diverse patient populations.
Add monitoring dashboards for hospital administrators to track triage efficiency and patient satisfaction metrics.