TRIAGE-IQ

Inspiration

I am a Nurse Manager of a Urology Unit at a Federal Teaching Hospital in Nigeria. Every day, my team faces a triage problem that no paper-based or spreadsheet system has solved well: Patients arrive in no particular order of clinical urgency. A man with suspected urosepsis might be sitting in the waiting area behind someone with a routine follow-up. A 3-hour-old testicular torsion — a surgical emergency with a hard 6-hour viability window — could be missed by a junior nurse who doesn't yet recognie the presentation. The consequences of poor triage order in urology are severe and sometimes irreversible: testicular loss from delayed torsion surgery; septic shock from untreated urosepsis; bladder rupture from missed acute retention. TRIAGE-IQ gives every nurse, regardless of experience level, the analytical depth of a senior clinician at the intake desk — instantly, consistently, and with documented reasoning. This isn't a speculative use case. It is my unit. These are my patients.

What it does

TRIAGE-IQ is a multi-step autonomous clinical triage agent built for Urology units in African federal teaching hospitals. When a patient arrives at the unit, a nurse enters the patient's presenting symptoms, chief complaint, and vital signs through a structured intake form. The agent then: Queries MongoDB Atlas (via MCP) to retrieve the patient's prior admission history and any flagged comorbidities stored in the unit's patient database Runs a multi-step clinical reasoning workflow powered by Gemini 3 — scoring symptom severity, cross-referencing known red-flag patterns (urosepsis triad, testicular torsion time window, acute urinary retention, AKI indicators), and computing a weighted triage priority Outputs a live priority queue — sorting all active patients by acuity, not just chronological arrival Generates a structured nursing handover note — a formatted clinical document ready for shift handover or escalation to the medical officer, including recommended immediate actions specific to the triage finding The agent moves beyond answering questions. It reads patient data, reasons over it, writes back a structured output, and manages a persistent queue — all under nurse oversight.

How we built it

Tech Stack: Gemini 3 — the clinical reasoning engine; runs multi-step symptom analysis and generates natural-language reasoning and handover notes Google Cloud Agent Builder — low-code agent orchestration layer, connecting the intake form to the reasoning pipeline and output generation MongoDB Atlas + MCP Server — patient records database; the MCP integration allows the agent to query prior visit history, comorbidities, and previous triage scores in real-time during the reasoning step Static HTML/JS frontend — single-file demo, deployable to any static host (Netlify Drop / Vercel) Agent Workflow (Multi-Step):

Patient Intake Submitted
        ↓
[Step 1] MongoDB MCP Query
        → Retrieve prior visits, known conditions, flagged history
        ↓
[Step 2] Gemini 3 Reasoning
        → Score symptoms against urology emergency criteria
        → Apply SIRS/sepsis screen
        → Check torsion time window
        → Weight vitals (HR, temp, SpO₂, BP)
        ↓
[Step 3] Priority Classification
        → IMMEDIATE / URGENT / ROUTINE
        → Red flag extraction
        ↓
[Step 4] Queue Reordering
        → Insert patient at correct position in live queue
        ↓
[Step 5] Handover Note Generation
        → Structured clinical document output
        → Written to MongoDB for shift records

Challenges we ran into

Clinical accuracy vs. completeness. Building a scoring system that catches true emergencies without over-triaging routine cases required careful calibration. I drew on established triage frameworks (Manchester Triage System, SIRS criteria) and validated the logic against real case presentations from my unit. Making a real-world agent feel genuinely useful, not theoretical. Many hackathon healthcare projects are demos for demo's sake. I constrained every feature to what a nurse could actually use in a 2-minute intake window under shift pressure. Every field in the form exists because removing it would reduce clinical accuracy. Balancing agent autonomy with nurse oversight. The agent prioritises and recommends — it never decides. Every output is a decision-support artefact that a nurse reviews and acts on. This is the right model for clinical AI at the current moment.

Accomplishments that we're proud of

Built a complete, functional triage agent grounded in real clinical criteria — not a generic "health chatbot" The handover note generator produces structured, ward-ready clinical documentation that maps directly to nursing handover standards at Nigerian federal teaching hospitals Four demo patients pre-loaded that illustrate the full triage spectrum: urosepsis, testicular torsion, acute urinary retention, and routine UTI — each handled correctly by the agent The project is built by someone who will actually use it. That is rare in healthcare AI hackathons.

What we learned

Google Cloud Agent Builder is genuinely powerful for rapid clinical workflow prototyping — the data store grounding feature is particularly valuable for connecting agents to existing medical record systems The MongoDB MCP integration was the architectural piece that transformed this from a standalone tool into an agent: the ability to pull prior patient context during the reasoning step is what makes the output clinically meaningful, not just algorithmically correct Clinical AI at the ward level doesn't need to be complex to be impactful. Triage ordering and documentation are unglamorous problems that cause real harm when done poorly.

What's next for TRIAGE-IQ

Expand to additional specialties — general surgery, paediatrics, emergency medicine; the core agent architecture is specialty-agnostic EMR integration — connect to Nigeria's DHIS2 health data infrastructure via API Voice intake mode — nurse dictates presenting complaint, Gemini transcribes and scores Shift analytics dashboard — trends in triage category distribution, peak admission windows, common red-flag presentations Multi-unit deployment — hospital-wide triage coordination, not just single-unit

Built With

• css
• gemini3
• googlecloudagentbuilder
• html
• javascript
• mongodbatlas
• mongodbmcpserver
• vercel