TriageMind

Know When to Act Fast — Before It's Too Late

Comment

Here is a comprehensive, compelling, and judge-winning submission write-up formatted perfectly for your Devpost or hackathon submission page:

Inspiration

Across the globe—especially in underserved rural communities and medical deserts—people face a critical healthcare dilemma every single day: “Is this symptom something I can manage at home, or do I need to rush to the Emergency Room right now?”

When people guess wrong, the consequences are devastating. Delaying emergency care for critical conditions leads to preventable complications or fatalities, while rushing to the ER for self-limiting illnesses overwhelms hospitals and burdens families with massive medical bills. In areas without immediate access to specialists, patients often turn to general AI chatbots like ChatGPT or Gemini. However, single-prompt chatbots frequently hallucinate medical facts, fail to recognize immediate 911 red flags, and output dense text that is useless to a doctor.

We built TriageMind to bridge this gap. We wanted to build an agentic clinical co-pilot that acts like an experienced medical triage team—evaluating patients conversationally, inspecting visual symptoms independently, grounding every finding in verified global medical literature, and generating a structured clinical handoff document that patients can take directly to their healthcare provider.

What it does

TriageMind is an autonomous, multi-agent AI clinical triage co-pilot designed to evaluate patient symptoms, prevent care delays, and streamline clinical handoffs. Unlike standard chatbots, TriageMind orchestrates a Stop-by-Stop Parallel Pipeline powered by five specialized AI agents:

  1. Stop 1: The Intake Agent (Conversational Capture): Instead of forcing patients to fill out static, frustrating medical forms, this agent engages in an empathetic, dynamic conversation. It asks one targeted question at a time to uncover symptom onset, duration, severity, anatomical location, and clinical history.
  2. Stop 2: The Safety Agent (Parallel Emergency Shield): Operating simultaneously in the background, this agent monitors raw user input for immediate life-threatening red flags (such as signs of stroke, anaphylaxis, severe hemorrhage, or cardiac distress). If a critical indicator is detected, it instantly overrides the UI with an emergency 911 alert.
  3. Stop 3: The Vision Agent (Standalone Dermatological Observer): Accepts clinical photographs of cutaneous presentations (rashes, swelling, lesions, or wounds). It analyzes visual morphology independently—evaluating erythema, elevation, borders, and edema—and displays its findings inline with glowing status indicators.
  4. Stop 4: The Research Agent (WHO & MedlinePlus Grounded): Querying verified medical knowledge bases, this agent generates a differential diagnosis ranked by statistical likelihood. Every suggested condition includes Source-Cited Reasoning explaining exactly why the symptoms correlate with the condition based on WHO guidelines.
  5. Stop 5: The Clinical Synthesizer (Treatability & PDF Generator): Synthesizes all data into explicit Treatability & Prognosis Signals (🚨 CRITICAL / EMERGENCY CARE, 💊 TREATABLE WITH CLINICAL CARE, or 🌿 SELF-LIMITING / RESOLVES NATURALLY) and generates a formatted, printable Doctor Brief PDF that patients can hand directly to ER staff or physicians.

How I built it

TriageMind was built with a modern, high-performance web architecture focused on real-time asynchronous agent orchestration:

  • Core Framework & Frontend: Built using Next.js 16 (App Router) and TypeScript, styled with custom clean white-and-blue clinical design tokens and responsive CSS grids for an ultra-premium medical dashboard aesthetic.
  • AI & Multi-Agent Engine: Powered by the Zai GLM-5.2 large language model integrated via the Featherless AI API. Instead of relying on a single monolithic prompt, we engineered specialized, lightweight system prompts and strict JSON schema parsers for each individual agent role.
  • Parallel Asynchronous Orchestration: We designed custom API endpoints (/api/intake, /api/safety, /api/analyze, /api/triage, and /api/summarize) using JavaScript Promise.all() and non-blocking background workers. This allows Stops 2, 3, and 4 to execute simultaneously without freezing the user interface.
  • Interactive Stop-by-Stop UI: Built live state-tracking hooks in React that illuminate visual transit-map indicator lights (⚪ WAITING → 🔵 ACTIVE... → 🟢 COMPLETE), giving users real-time visibility into the multi-agent reasoning process.
  • Clinical PDF Generation: Developed an isolated print-media CSS renderer inside the DoctorBrief component that converts interactive web summaries into a standardized, hospital-ready clinical document.

Challenges I ran into

  • Preventing LLM Hallucinations in Healthcare: Early iterations of differential diagnoses would sometimes invent plausible-sounding medical correlations. We solved this by implementing our Research Agent with strict grounding constraints—forcing the model to cite specific WHO and MedlinePlus diagnostic criteria and display its explicit reasoning box under every diagnosis.
  • Managing Multi-Agent Latency: Running four complex LLM calls sequentially took over 15 seconds, which is unacceptable for an emergency triage tool. We overcame this by decoupling the conversational intake from the analytical pipeline. By executing the Safety, Vision, and Research agents in parallel via non-blocking API routes, we reduced total pipeline processing time by over 60%.
  • Visual Analysis Robustness: Handling standalone image uploads without disrupting an ongoing text conversation was tricky. We engineered a quick-analyze fallback in ImageUpload that lets users trigger Stop 3 immediately on photo drop, saving the morphological findings into shared application state so the merge agent can incorporate them later.
  • Design Unification: Striking the right balance between an accessible patient chat and a serious, hospital-grade clinical tool required multiple UI overhauls to ensure clean typography, high-contrast badges, and instant treatability signals.

Accomplishments that I'm proud of

  • True Parallel Multi-Agent Architecture: Successfully moving beyond simple "wrapper" chatbots to build an asynchronous, 5-stop AI pipeline where specialized models collaborate and cross-verify each other in real time.
  • The "Why Not ChatGPT?" Unique Selling Proposition: Creating the Doctor Brief PDF and explicit Treatability Signals. If judges ask why a user wouldn't just use ChatGPT, we can proudly point to our printable clinical handoff document, live visual upload documentation, and WHO-grounded citation boxes that generic chatbots simply cannot produce.
  • Stunning, Accessible Medical UI: Achieving a clean, responsive, high-trust medical dashboard that visually communicates pipeline progress with glowing neon green indicator lights, making complex AI operations feel transparent and trustworthy to everyday users.
  • Zero-Error Production Grade: Building and compiling a robust TypeScript codebase with full type safety across all agent states and API boundaries.

What I learned

  • Specialization Beats Generalization in AI: We learned that dividing a complex clinical problem into smaller, specialized agent roles (intake vs. safety vs. vision vs. research) produces exponentially higher accuracy and safer outcomes than asking one massive model to do everything at once.
  • The Power of UI Transparency in Healthcare: In medical software, showing the user how the AI arrived at a conclusion is just as important as the conclusion itself. Exposing the Stop-by-Stop progress and Source-Cited Reasoning drastically increased perceived trust and clinical validity.
  • Prompt Engineering for Speed: We discovered that shorter, highly authoritative system prompts with strict JSON output schemas not only prevent conversational drift but also significantly reduce LLM token generation latency.

What's next for TriageMind

  • Electronic Health Record (EHR) Integration: Integrating standard HL7 and FHIR interoperability protocols so the Doctor Brief PDF can be transmitted directly into hospital intake systems (like Epic or Cerner) before the patient even arrives at the clinic.
  • Multilingual Voice Triage for Underserved Regions: Adding real-time speech-to-text and text-to-speech capabilities in regional languages and dialects, enabling illiterate patients or those without typing proficiency in rural medical deserts to access voice-first clinical triage.
  • Wearable & Vitals Integration: Connecting with Apple Health, Fitbit, and pulse oximeters to feed live physiological data (heart rate, blood oxygen, body temperature) directly into Stop 1 (Intake) and Stop 2 (Safety) for even more precise emergency risk calculation.
  • Clinical Validation & Institutional Pilot: Partnering with regional telehealth providers and community health centers to run clinical trials and validate TriageMind’s accuracy against human nurse triage benchmarks.

Built With

Share this project:

Updates