RawatAI — Malaysia Public Healthcare Agentic AI

An AI-powered patient triage and hospital routing system built for the Adaptive Malaysia Hackathon 2026. Patients describe symptoms, upload a photo, and a 4-agent autonomous pipeline classifies urgency, predicts the likely condition, finds the nearest appropriate facility, and assigns a specialist — all without human intervention.

---

Quick Start

1. Add your Groq API key

Create .env.local in the project root:

GROQ_API_KEY=your-groq-api-key-here

Get a free key at https://console.groq.com

2. Run

npm install
npm run dev

Open http://localhost:3000

---

Demo Accounts

Name	Phone	Password	Conditions
Ahmad Razif	0123456789	demo123	Hypertension
Siti Nurhaliza	0198765432	demo123	Type 2 Diabetes

Doctor accounts have been removed. The system is fully autonomous — no doctor login required.

---

Demo Flow

1. Visit http://localhost:3000 → click Start Triage →
2. Login with a demo patient account
3. On the AI Health Assessment page:
   • Type or speak your symptoms (voice input via Web Speech API)
   • Optionally upload a photo (injury, rash, prescription)
   • Select your city from the dropdown or use GPS
   • Toggle EN / BM for Bahasa Malaysia output
4. Click Run AI Triage — watch 4 agents process in real time:
   • Agent 1 → Vision (image analysis)
   • Agent 2 → Triage (urgency classification + disease prediction)
   • Agent 3 → Routing (nearest facility based on live capacity)
   • Agent 4 → Assignment (specialist matching)
5. Review results:
   • Priority badge — P1 Emergency / P2 Urgent / P3 Non-urgent
   • Predicted condition with department label and key symptoms
   • Assigned specialist card — Call Hospital button, Book Appointment with crowd bar
   • Recommended hospital with map, distance, and occupancy
   • Surge auto-reroute — if top hospital is CRITICAL, system auto-selects next best
6. Expand AI Agent Technical Details at the bottom to see the full agent reasoning log

---

Agentic Pipeline

Patient Input (symptoms + photo + location)
        │
        ▼
┌─────────────────────────────────────────────────────┐
│  Agent 1 — Vision Agent                             │
│  Model: LLaMA 4 Scout (multimodal)                  │
│  Input: base64 image + symptom text                 │
│  Output: clinical findings from image               │
└───────────────────┬─────────────────────────────────┘
                    │
                    ▼
┌─────────────────────────────────────────────────────┐
│  Agent 2 — Triage Agent                             │
│  Model: LLaMA 3.1-8B-Instant (Groq)                 │
│  Input: symptoms + vision findings + medical history│
│  Output: priority (P1/P2/P3), predicted_disease,    │
│          department, key_symptoms, reasoning_steps  │
│  Rules: fever >2 days → P2 min; hypertension +      │
│         fever → P2 min; diabetes + fever → P2 min  │
└───────────────────┬─────────────────────────────────┘
                    │
                    ▼
┌─────────────────────────────────────────────────────┐
│  Agent 3 — Routing Agent                            │
│  Input: lat/lon + priority + ICU flag               │
│  Output: ranked facilities (Haversine + utilization)│
│  Data: 3,304 MoH facilities + 149 bed util rates   │
│  Surge: if top facility CRITICAL → auto re-route    │
└───────────────────┬─────────────────────────────────┘
                    │
                    ▼
┌─────────────────────────────────────────────────────┐
│  Agent 4 — Assignment Agent                         │
│  Input: priority + key_symptoms + location          │
│  Output: matched specialist from virtual pool       │
│  Pool: 16 doctors, 9 specialties, across Malaysia  │
└───────────────────┬─────────────────────────────────┘
                    │
                    ▼
         Patient sees results + books appointment

---

Key Features

Feature	Details
Voice input	Web Speech API — speak symptoms in EN or BM
Bilingual	EN / BM toggle — AI output switches language
Disease prediction	AI returns predicted_disease + department
Surge detection	Auto-reroutes if top hospital is CRITICAL capacity
Book appointment	One-click booking with live hospital crowd bar
Patient-friendly loading	Plain-language progress steps, technical log hidden
Medical history context	Chronic conditions + medications fed into triage prompt

---

Project Structure

rawatai/   (project root)
├── app/
│   ├── page.tsx                  — Landing page
│   ├── login/page.tsx            — Patient login
│   ├── patient/
│   │   ├── dashboard/page.tsx    — Patient dashboard
│   │   ├── triage/page.tsx       — Main AI assessment page
│   │   └── history/page.tsx      — Past assessments
│   └── api/
│       ├── triage/route.ts       — LLaMA 3.1 triage + disease prediction
│       ├── vision/route.ts       — LLaMA 4 Scout image analysis
│       ├── routing/route.ts      — Facility scoring + surge detection
│       ├── assign/route.ts       — Specialist assignment
│       └── notify/route.ts       — Case persistence (in-memory)
├── lib/
│   ├── routing.ts                — Haversine distance + facility scoring
│   ├── doctors.ts                — Virtual specialist pool (16 doctors)
│   ├── locations.ts              — 29 Malaysian city coordinates
│   ├── session-store.ts          — globalThis in-memory store (HMR-safe)
│   └── demo-users.ts             — Demo patient accounts
├── components/
│   ├── PatientNav.tsx            — Sidebar navigation
│   └── FacilityMap.tsx           — Leaflet map component
└── public/
    └── logo.png                  — Malaysia coat of arms

---

Datasets

Hosting note: The frontend/ folder is fully self-contained. All facility data has been pre-processed from the raw CSVs into frontend/lib/facilities.json, which is bundled with the Next.js app at build time. The /datasets folder contains the original source files for reference only — it is not required at runtime or for deployment.

File	Location	Description
facilities.json	lib/	Pre-baked from CSVs — used by the app
facilities_master.csv	datasets/	3,304 Malaysian government health facilities with GPS (source)
bedutil_facility.csv	datasets/	149 hospital bed utilization rates — MoH Malaysia (source)
bedutil_state.csv	datasets/	State-level capacity baselines (source)

Source: Ministry of Health Malaysia open data

---

Tech Stack

Layer	Technology
Framework	Next.js (App Router) + React 19 + TypeScript
Styling	Tailwind CSS v4 + inline styles
AI — Triage	Groq API · LLaMA 3.1-8B-Instant
AI — Vision	Groq API · LLaMA 4 Scout (multimodal)
Routing	Haversine distance + MoH bed utilization data
State	In-memory Maps anchored to globalThis (no database)
Icons	react-icons/ri
Maps	Leaflet (via react-leaflet)

---

Architecture Notes

• Self-contained for deployment — only the frontend/ directory needs to be deployed. All facility data is pre-baked into frontend/lib/facilities.json. The /datasets and /agent-service folders are not needed at runtime.
• No database required — all session state is held in globalThis Maps that survive Next.js HMR reloads in development.
• No doctor interface — the system is fully autonomous. The assignment agent replaces the manual doctor confirmation loop.
• Bilingual AI — the triage system prompt switches language based on the lang parameter (en / bm).
• Surge resilience — if the routing agent's top recommendation has CRITICAL capacity, it automatically promotes the next-best facility and shows a warning banner.

Deploying to Vercel

vercel --prod

Add GROQ_API_KEY as an environment variable in your Vercel project settings. No root directory override needed — the Next.js app is at the repo root.