AI SMART CITY

Backend API docs
Nova AI Output
Frontend Homepage
Report Page with Nova Result
Dashboard
GitHub repo

Inspiration

Every day, city staff manually read hundreds of citizen complaints — potholes, broken streetlights, flooded roads. It's slow, error-prone, and doesn't scale. I built this to automate the entire triage pipeline using Amazon Nova AI.

Community Impact

Urban infrastructure problems affect millions of citizens daily — but most cities are overwhelmed by the volume of complaints and understaffed for manual triage.

This project creates real, measurable impact at three levels:

For citizens: Reporting a problem takes 10 seconds — one photo, one sentence. No navigating complex government portals, no waiting on hold, no wondering if anyone saw your complaint.

For city staff: Zero manual triage. Every complaint arrives pre-classified, pre-routed, and pre-described by Nova. A team that previously spent 3 hours/day reading complaints can now focus entirely on fixing them.

For city managers: Nova's cluster summaries automatically surface patterns — 8 flooding complaints on the same street, recurring power outages in one area — that would take hours to spot manually. This enables proactive infrastructure maintenance instead of reactive firefighting.

Scale: Every city in the world faces this problem. This API is stateless, rate-limited, and production-ready — it can plug into any municipality's existing workflow today.

How I built it

Backend: FastAPI + Uvicorn
AI: Amazon Nova Lite (multimodal) via AWS Bedrock (boto3)
Database: PostgreSQL with full complaint + status history
Frontend: Multimodal UI for citizen reporting
Nova powers 4 features: Vision Analysis, Intelligent Classification, Conversational Intake, and Cluster Summaries for managers

Challenges

Getting Nova to return consistent structured JSON for classification
Handling multimodal inputs (image + text) together in a single API call
Building conversational intake that guides citizens naturally to provide all required info

What I learned

Nova's vision + text combination produces better classification than either alone. Confidence scores are genuinely meaningful when the image backs up the text.