EnvRoute

Real-time flood detection + AI-powered routing to keep cities moving when it rains.

Comment

Key capabilities:

  • 🔍 AI Flood Detection — YOLOv10 model trained on 7,400+ flood/non-flood images analyzes live camera feeds with confidence scoring
  • 🗺️ Real-Time Flood Risk Map — Interactive map with color-coded severity levels (safe / warning / flooded / severe)
  • 🧭 Smart Routing — Multi-factor optimization balancing flood risk, travel time, and distance
  • 🤖 Multi-Agent Intelligence — Autonomous agents for data crawling, pattern analysis, and decision-making
  • 💬 LLM-Powered Reports — Natural-language flood situation summaries and travel advisories
  • 🔔 Alert System — Push notifications and email alerts for flood events on your routes

## How we built it

We designed EnvRoute as a microservices architecture with four major layers:

### AI/ML Layer

  • Trained a YOLOv10 flood detection model using transfer learning on 7,400+ curated images
  • Used ClimateGAN for synthetic data augmentation to improve model robustness across varied weather conditions
  • Built a multi-agent system with LangChain — data crawling agents ingest camera/sensor/weather feeds, analysis agents detect patterns, and decision agents optimize routes and dispatch alerts

### Backend (Python + FastAPI)

  • 7 independent microservices: auth, camera ingestion, flood detection, flood information, geospatial tracking, real-time monitoring, and notifications
  • Apache Kafka for event-driven streaming between services
  • Redis for real-time state and pub/sub
  • PostgreSQL + Supabase for persistent storage
  • All orchestrated via Docker Compose

### Mobile App (Flutter)

  • Cross-platform iOS/Android app with Google Maps SDK
  • Real-time flood-aware navigation with dynamic rerouting
  • BLoC pattern for state management
  • Firebase Auth + Supabase for authentication and real-time data sync

### Web Dashboard (React)

  • React 18 + TypeScript with Ant Design
  • Leaflet maps for flood monitoring and camera feed visualization
  • Admin panel for analytics, camera management, and alert configuration

### Landing Page

  • React 19 + Vite + TailwindCSS 4 + Framer Motion for a polished, animated presentation at envroute.skylabs.vn

## Challenges we ran into

  • Limited real-world flood imagery — Flood events are sporadic and hard to capture. We overcame this by using ClimateGAN to generate synthetic flood scenes, significantly expanding our training dataset
  • Real-time inference latency — Processing live camera feeds through YOLOv10 while maintaining sub-second detection required careful optimization of our inference pipeline
  • Microservices coordination — Keeping 7 services in sync with consistent data flow across Kafka topics was complex, especially handling edge cases like camera disconnections or delayed sensor data
  • Dynamic routing under uncertainty — Building a routing engine that balances multiple objectives (flood risk vs. travel time vs. distance) required iterative tuning of our cost function
  • Multi-agent conflict resolution — Ensuring crawling, analysis, and decision agents collaborate without contradictory outputs demanded careful orchestration design

## Accomplishments that we're proud of

  • End-to-end working pipeline: camera feed → AI detection → risk map update → user rerouting → alert notification
  • Multi-platform delivery in 48 hours: native mobile app, web dashboard, microservices backend, AI model, and landing page
  • YOLOv10 flood detection achieving reliable performance with confidence-scored predictions
  • LLM-powered intelligence generating human-readable flood reports from raw sensor data
  • Live deployment accessible at envroute.skylabs.vn

## What we learned

  • Training and deploying computer vision models for real-world environmental monitoring
  • Designing event-driven microservices with Kafka and FastAPI at speed
  • Building multi-agent AI systems that coordinate data collection, analysis, and autonomous decision-making
  • The complexity of real-time geospatial processing and dynamic route optimization
  • How to ship a full-stack, multi-platform product under extreme time pressure

## What's next for EnvRoute

  • Expand camera coverage to all flood-prone areas across Ho Chi Minh City
  • Add flood prediction using weather forecasts, historical patterns, and tidal data
  • Integrate with city traffic management systems for coordinated emergency response
  • Partner with ride-hailing platforms (Grab, Be) to provide flood-aware routing at scale
  • Deploy edge inference on traffic cameras for faster, decentralized detection
  • Add community-sourced flood reports to complement camera-based detection

Built With

Share this project:

Updates