City Pulse

Citizens snap a photo of a city issue — AI classifies it instantly, clusters hotspots on a live map, and generates council briefings. Every phone becomes a smart city sensor.

Comment

Inspiration

Cities deploy thousands of expensive IoT sensors to monitor infrastructure — but every citizen already carries a powerful sensor in their pocket. I asked: what if we could turn every smartphone into a smart city sensor? CityPulse was born from the idea that civic reporting should be as easy as taking a photo.

What it does

Citizens snap a photo of an urban issue — pothole, broken streetlight, graffiti, flooding — and CityPulse handles the rest:

  • AI classifies the issue type, severity, and responsible department in seconds (Groq Llama 4 Scout vision)
  • DBSCAN clusters nearby reports into actionable hotspots
  • Interactive dashboard shows a live map with severity-colored markers, heatmap density overlay, and city health metrics
  • AI chat assistant answers questions about city conditions using report data and live local news
  • City council briefings are auto-generated — professional memos ready for officials
  • Real-time updates via SSE — submit on your phone, watch it appear on the dashboard instantly
  • Citizens can upvote/verify reports, and severity auto-escalates after 3 confirmations
  • Reports flow through a resolution workflow (open → in progress → resolved)
  • Multi-city support — add any city with a config entry
  • Privacy-first — all photo metadata (EXIF/GPS) is stripped before storage
  • Voice reporting via Web Speech API for accessibility
  • Accessibility impact score weights issues by their effect on disabled citizens
  • Neighborhood risk scoring grades areas by report density, severity, and resolution rates
  • GeoJSON open data API for integration with external GIS tools

How I built it

  • Backend: FastAPI with SQLite/SQLAlchemy — simple, fast, no infrastructure overhead
  • AI: Groq API — Llama 4 Scout for vision classification, Llama 3.1 8B for chat, briefings, and news translation
  • Clustering: scikit-learn DBSCAN on raw lat/lng coordinates (eps=0.003 ≈ 300m)
  • Map: Folium/Leaflet with heatmap overlay, LayerControl for toggling layers
  • Real-time: Server-Sent Events with asyncio.Queue per connected client
  • News: SWR RSS feeds filtered by city keywords, auto-translated to English via Groq
  • Seed data: 40 real street-level photos from Mapillary API, each AI-classified
  • Privacy: Pillow strips all EXIF metadata from uploaded photos
  • Frontend: Vanilla HTML/CSS/JS — no frameworks, dark mode, PWA-ready
  • Testing: 47 pytest tests covering classification, clustering, health score, submissions, and dashboard
  • Deployment: VPS with nginx + Let's Encrypt SSL at citypulse.help

Challenges I ran into

  • Gemini API quota exhaustion — Google's free tier ran out mid-development. I pivoted to Groq (Llama 4 Scout) which has a more generous free tier and faster inference
  • RSS news quality — German regional news feeds include stories from the entire state. I built a keyword filter for city-specific news and auto-translate German headlines to English via Groq
  • Folium limitations — Folium generates static HTML, so I can't dynamically add markers. I solved real-time updates with SSE toast notifications instead
  • Street-level photos — Mapillary coverage is patchy. I implemented progressive search radius (300m → 600m → 1km) to maximize coverage, getting 40/50 locations covered

Accomplishments I'm proud of

  • 22+ features built and deployed solo — every one peer-reviewed by adversarial AI agents
  • Zero unclassified reports — every seed photo was AI-classified, with Groq finishing what Gemini couldn't
  • 47 passing tests with zero regressions throughout development
  • Real Mapillary photos from actual Stuttgart streets — not stock photos or placeholders
  • Multi-model AI architecture — vision classification, conversational chat, council briefings, and news translation all running on Groq's free tier
  • The app is live and functional at citypulse.help — not a prototype

What I learned

  • Free AI APIs have real quota limits — always have a fallback provider
  • Adversarial peer review (even between AI agents) catches bugs that solo development misses
  • For hackathons, a working deployed product beats a feature list every time
  • Folium is great for quick map prototyping but has limitations for real-time interactivity
  • Privacy features (EXIF stripping) are trivial to implement but powerful for user trust

What's next for CityPulse

  • More cities — the multi-city architecture is ready, just needs data seeding
  • User accounts — track your reports, get notifications when they're resolved
  • Department portal — city officials manage assignments and SLA tracking
  • Before/after photos — citizens document when issues are fixed
  • Predictive maintenance — use historical patterns to schedule preventive repairs
  • Native mobile app — PWA works today, but a native app could access camera and GPS more seamlessly

Built With

Share this project:

Updates