UofTHacks-2026

TriageAI

Emergency departments are often overwhelmed, and the first few minutes of intake can significantly impact patient outcomes. Today, triage relies heavily on manual questioning, language-dependent communication, and subjective interpretation—often under stressful conditions.

We were inspired by a simple question: what if we could preserve a patient’s story, context, and urgency clearly and safely before a clinician ever sees them?

TriageAI was built to explore how responsible AI, when designed with guardrails and clinicians in the loop, can improve accessibility, consistency, and safety in emergency care.

What it does

TriageAI is an AI-assisted, clinician-in-the-loop triage system designed for emergency rooms and urgent care settings.

It consists of:

• A patient-facing web intake kiosk where patients describe their symptoms using speech or text
• An AI-powered multi-agent pipeline that structures intake, estimates urgency, and produces a clinician-ready summary
• A doctor dashboard that shows patients prioritized by urgency and clearly visualizes how AI agents contributed to each assessment

Key features:

• Multilingual-friendly intake via speech or text
• Passive, non-invasive vitals support (optional)
• Conservative urgency classification (Low / Medium / High / Critical)
• Deterministic safety guardrails that always err on the side of escalation
• Clear, explainable agent hand-offs and timestamps
• No diagnoses or treatment recommendations — decision support only

How we built it

TriageAI was built as a monorepo with a clean separation of concerns:

• Frontend

  • React + TypeScript + Tailwind CSS
  • Patient Intake Web App (ER kiosk)
  • Doctor Dashboard Web App (clinician-facing)

• Backend

  • Node.js + Express API
  • MongoDB as a shared, evolving patient state store
  • Polling-based real-time updates for demo reliability

• AI & Orchestration

  • Multi-agent system orchestrated using LangGraph
  • Agents include:
    • Intake Structuring Agent
    • Urgency Classification Agent
    • Clinician Summary Agent
  • Deterministic post-agent guardrails enforce safety rules (age risk, red flags, vitals, uncertainty)
  • Adaptive memory and model selection supported via Backboard.io + LangChain

The MongoDB patient record acts as the shared “state,” allowing each agent to append its output in sequence, making orchestration and hand-offs fully inspectable.

Challenges we ran into

• Balancing safety with usefulness in a healthcare context
  We had to be extremely careful to avoid diagnoses, treatment advice, or overconfident AI outputs.

• Making multi-agent orchestration visible
  It’s easy to claim “multi-agent AI”; it’s much harder to show it clearly. We solved this by explicitly exposing agent timelines, outputs, and timestamps in the dashboard.

• Scope discipline as a solo hacker
  Many tempting features (real-time video analysis, mobile apps, advanced analytics) were intentionally cut to ensure a stable, polished demo.

Accomplishments that we’re proud of

• Designing a responsible AI system for healthcare that prioritizes safety and explainability
• Building a clear, inspectable multi-agent pipeline that meets real-world standards
• Creating a clinician dashboard that makes AI behavior transparent rather than opaque
• Integrating deterministic guardrails that constrain AI outputs
• Completing an end-to-end system solo within a tight hackathon timeframe

What we learned

• In high-stakes domains, constraints are a feature, not a limitation
• AI systems are far more trustworthy when paired with explicit rules and human oversight
• Clear state management and explainability matter more than model complexity
• Judges and users alike value clarity, responsibility, and intent over flashy demos

What’s next for TriageAI

• A home/mobile pre-triage experience to help patients choose the right care setting before arriving
• Deeper multilingual support with omnilingual speech models
• Integration with EHR systems for seamless clinician workflows
• Longitudinal learning from clinician feedback to continuously improve triage consistency
• Further research into bias mitigation and accessibility in emergency care

Running the Demo

Prerequisites

• Node.js 20.6+ (uses node --env-file)
• A .env file in the repo root:

mongodb_cluster_username=...
mongodb_cluster_password=...
backboard_api_key=...

If MongoDB Atlas is unreachable (e.g. your IP isn't on the cluster's Network Access list), the server falls back to in-memory storage so the demo still works. If Backboard.io is unreachable, agents fall back to deterministic logic. Both fallbacks log a warning.

Start all three services

# Terminal 1 — backend API (http://localhost:3001)
cd server && npm install && npm run dev

# Terminal 2 — patient intake kiosk (http://localhost:3000)
cd apps/patient-intake && npm install && npm run dev

# Terminal 3 — doctor dashboard (http://localhost:3002)
cd apps/doctor-dashboard && npm install && npm run dev

Optional: seed demo patients so the dashboard isn't empty:

cd server && npm run seed

Demo script

1. Open the dashboard (:3002) on one screen and the kiosk (:3000) on another.
2. Complete an intake on the kiosk — try mentioning "chest pain" and an age over 65 to see red flags and guardrail escalations.
3. Watch the dashboard: the new patient appears within 5 seconds, and the detail view shows each agent going pending → running → completed live.
4. Review the final urgency, applied guardrails, and clinician summary.
5. Use "Clear Demo Data" in the dashboard nav to reset between runs.