FloodSense

8.9 million motorcycles. Zero flood warnings. FloodSense uses AI, live cameras, and real-time weather to tell you which roads are underwater — before you ride into them.

Comment

Inspiration

Every rainy season in Ho Chi Minh City, 8.9 million motorcycles face the same gamble: which roads are flooded right now? Google Maps knows the fastest route — but it has zero idea which streets are underwater. We've both been that rider, stuck knee-deep on Nguyen Huu Canh at 6pm, watching our commute turn into a 2-hour nightmare.

HCMC averages 15–30 flood points per heavy rain, 3–4 times a week during the 6-month monsoon season. Flood information exists — scattered across Facebook groups, news sites, and government bulletins — but none of it is structured, mapped, or actionable in real time. We built FloodSense to fix that.

What it does

FloodSense is a real-time flood intelligence map for HCMC that fuses multiple signal layers into one actionable dashboard:

  • Live Flood Heatmap — Color-coded flood zones (red = heavy >30cm, orange = moderate, yellow = light) on a Mapbox-powered map, auto-refreshing every 60 seconds
  • AI Intelligence Pipeline — Exa.ai neural search scrapes Vietnamese news and social media every 5 minutes; GPT-4o extracts structured flood data (street, district, depth, confidence) from articles
  • Route Flood Check — Enter origin → destination, and FloodSense highlights which segments of your route intersect active flood zones, with Vietnamese voice warnings
  • Live Camera Monitoring — Public traffic cameras at known flood hotspots provide visual confirmation of road conditions, feeding flood signals back into the confidence scoring system
  • Proactive Alerts — Save your daily commute routes; when a new flood is detected on your path, you get a push notification + Vietnamese voice alert before you leave home
  • Voice Alerts — ElevenLabs-powered natural Vietnamese TTS (Southern female / Northern male voice options) for hands-free flood warnings while riding
  • Crowdsource Reports — Riders can report floods directly; 2+ independent confirmations within 200m auto-promote a report to a verified flood event
  • Weather Pre-warning — 6-hour rainfall forecast from Open-Meteo + government alerts to predict flood risk before rain even starts

The confidence scoring uses signal fusion: weather data raises risk, news/social scraping raises evidence, camera feeds and user reports raise confirmation. No single source is trusted alone.

How we built it

Frontend: React 18 + TypeScript + Vite, styled with TailwindCSS. Mapbox GL JS handles the flood heatmap with custom layers and route overlays. State management via Zustand with persist middleware for offline-friendly saved routes.

Backend: Express.js on Node.js, deployed as a single Vercel serverless function. PostgreSQL on Supabase with PostGIS extensions for spatial queries — enabling operations like "find all floods within 200m of this route's bounding box" in \( O(1) \) using spatial indexes.

AI Pipeline: $$ \text{Exa.ai (neural search)} \xrightarrow{\text{5 queries}} \text{GPT-4o (extraction)} \xrightarrow{\text{dedup 300m}} \text{Supabase (upsert)} \xrightarrow{\text{notify}} \text{User} $$

Each flood event has a 2-hour TTL — if no new source confirms it, it auto-expires. Cross-validation requires \( \geq 2 \) independent sources for "High" confidence.

Camera Layer: Hybrid approach — real HCMC traffic camera locations at flood hotspots (Nguyen Huu Canh, Dinh Bo Linh, Huynh Tan Phat, etc.) with flood detection status signals that feed into the existing confidence scoring.

Automation: n8n orchestrates the background intelligence loop — cron every 5 minutes triggers the Exa→GPT-4o→Supabase pipeline, then checks saved routes for affected users and fires push notifications.

Key APIs: Exa.ai (neural search), OpenAI GPT-4o (extraction), ElevenLabs (Vietnamese TTS), Mapbox (maps + directions + geocoding), Open-Meteo (weather), OpenWeather (government alerts), Supabase (database + realtime).

Challenges we ran into

  1. Vietnamese NLP is hard — GPT-4o occasionally misparses Vietnamese street names with diacritics. "Dinh Bo Linh" vs "duong DBL" all refer to the same road. We implemented fuzzy street matching with PostgreSQL's pg_trgm GIN indexes to handle typo-tolerant deduplication.

  2. Vercel's 12-function limit — Our Express server initially generated too many serverless function entry points. We had to restructure the entire API into a single api/index.ts entry point with internal routing, prefixing helper modules with _ to exclude them from Vercel's function detection.

  3. Flood deduplication radius — Too small (100m) and the same flood appears as 3 separate events. Too large (500m) and two genuinely different floods on parallel streets merge into one. We landed on 300m for news/AI sources and 200m for crowdsource reports after testing with real HCMC street density.

  4. Demo reliability vs. real data — The app needs rain to show floods, but hackathon demos don't wait for weather. We built a "Simulate Rain" mode that rolls out 15–20 preset flood points with staggered animation, triggering the full notification + voice alert pipeline — so the demo always works regardless of actual weather.

  5. ElevenLabs latency — Initial voice generation took 4–5 seconds, breaking the "instant alert" feel. We implemented audio caching in the alert_history table — if the same alert text was already generated, serve the cached audio URL instead of calling the API again. Got it down to <1 second for repeated alerts.

What we learned

  • Signal fusion > single source — No individual data source (news, social media, weather, cameras) is reliable enough alone. The power is in combining weak signals into strong confidence scores.
  • Mock-first design pays off — Building every feature with a mock fallback meant we were never blocked by API outages or missing credentials. The VETC integration is mock, the camera detection layer uses simulated status — but the architecture is ready for real data.
  • Spatial databases are underrated — PostGIS turned "find floods near this route" from an \( O(n) \) loop into an indexed spatial query. For a map-heavy app, this was a game-changer.
  • Vietnamese voice UX matters — Offering both Southern and Northern Vietnamese voices wasn't just a feature checkbox — our test users immediately had a preference, and it made the alerts feel personal rather than robotic.

Built With

Share this project:

Updates