🌦️ Weatherise — Multi-Agent System for Travel Optimization in Da Nang City

Team: Weatherise (4 members) Hackathon: Vietnam AI Open Hackathon 2026

1. Project Overview
2. Why This Matters Now
3. Platform Strategy
4. System Architecture
5. Future Reuse & Enterprise Value

Weatherise is a domain-agnostic Multi-Agent System (MAS) that combines real-time weather intelligence with optimization algorithms to deliver hyper-personalized travel recommendations for Da Nang City. The system orchestrates specialized AI agents — each powered by GPU-accelerated models running on an 8× H200 cluster — to reason about weather forecasts, tourist attractions, route optimization, and local expertise simultaneously.

User Query → Orchestrator Agent → [Weather Agent + Attraction Agent + Route Agent + Local Expert Agent] → Optimized Travel Plan
                                                          ↕ (real-time loop)
                                          Weather Watcher Agent → SMS / WebSocket Alert → Dynamic Re-plan

The architecture is built as a reusable multi-agent framework — swap the domain knowledge (travel → agriculture, logistics, disaster response) and the same agent orchestration, RAG pipeline, and optimization layer works across industries.

• 2025: 17.3M visitors (+15% YoY), VND 60 trillion revenue (+21%)
• 2026 target: 19.5M visitors — first 5 months already at 7.74M (+20.9%)
• Da Nang named Vietnam's Smart City for 6 consecutive years
• City already deploying AI chatbots, AR experiences, and the "Danang Smart City" super-app

The problem: Tourists still manually check weather, browse blogs, and guess which attractions to visit. There is no intelligent system that combines weather awareness with real-time travel optimization.

• Weather is the #1 factor affecting tourist satisfaction (World Tourism Organization, UNWTO)
• 68% of travelers change plans due to unexpected weather (Booking.com Travel Report 2025)
• Climate volatility increasing — Da Nang faces typhoon season (Sep-Dec), extreme heat (Jun-Aug)

For a 5-day hackathon with 4 members, here's the optimal strategy:

Day 1-3 (MVP):     Backend MAS + Gradio/Streamlit Chat UI
Day 4-5 (Polish):  Next.js Responsive PWA with map integration

graph LR
    subgraph "Phase 1: MVP (Day 1-3)"
        A[Gradio Chat UI] --> B[FastAPI Backend]
        B --> C[Multi-Agent System]
    end
    subgraph "Phase 2: Polish (Day 4-5)"
        D[Next.js PWA] --> B
        D --> E[Leaflet Map]
        D --> F[Weather Dashboard]
    end
    style A fill:#4CAF50,color:#fff
    style D fill:#2196F3,color:#fff

Judge Strategy: Judges can open the web app on any device (laptop/phone). PWA feels like a native app. The Gradio fallback ensures the system is always demonstrable even if the frontend isn't ready.

• Gradio: MVP Chat UI
• Next.js 14: PWA Polish
• Leaflet.js: Maps
• Recharts: Weather Charts

• FastAPI: API Gateway
• WebSocket: Streaming
• Celery: Task Queue

• LangGraph: Orchestration
• LangChain: Agent Tools
• NIM: LLM Serving
• NeMo Guardrails: Safety
• APScheduler: Weather Watcher

• LLM Models (via NIM):
  • Nemotron-3 Super 49B
  • Llama 3.1 70B Instruct
• Embedding & Reranker:
  • NV-Embed-v2
  • NV-Rerank-Mistral-4B
• Weather Foundation Models:
  • Earth-2 Atlas (Medium Range 15-day Forecast)
  • Earth-2 FourCastNet (Fallback)
• High Resolution / Nowcasting (Future):
  • Earth-2 CorrDiff (Downscaling)
  • Earth-2 StormScope (Nowcasting)

• Milvus: Vectors
• PostgreSQL: Structured
• Redis: Cache + Sessions

• Docker + Docker Compose
• Nginx: Reverse Proxy
• Prometheus + Grafana: Monitoring
• JupyterLab: Dev Environment

• cuOpt: Route Optimization
• Earth2Studio: Weather Framework
• GFS: Initial Conditions
• NIM: Containers
• NeMo: Toolkit

• SpeedSMS: Vietnam SMS
• Open-Meteo: Free Weather
• Playwright: Plan Image
• qrcode: QR Generator