Civic Infrastructure AI
City-scale detection of 180+ asset types with 95%+ accuracy
The Problem
Manual inspection of street-level imagery for infrastructure assets at city scale. A major municipal government needed to catalog and monitor infrastructure assets — fire hydrants, street signs, utility poles, road markings, and 170+ other asset types — across an entire city. Manual inspection was slow, expensive, and could not scale. Street-level imagery existed but lacked automated analysis. PII in public imagery (faces, license plates) created compliance risk that blocked deployment.
Our Approach
We built a production AI pipeline that combines Vision AI for multi-class asset detection with LiDAR point cloud processing for precise geospatial mapping. The system runs on AWS ECS Fargate with 48 parallel workers to process city-scale imagery in production timeframes.
PII redaction was built as a first-class capability — every image passes through automated face and license plate detection before any analysis occurs. The entire pipeline is tested with 2,000+ automated tests and monitored in production with comprehensive alerting.
Architecture
From image ingestion to geospatial asset mapping
Image Ingestion
Street-level imagery collected at city scale, processed through a parallelized ingestion pipeline on ECS Fargate with 48 concurrent workers.
AI Detection Pipeline
Vision AI classifies 180+ infrastructure asset types from imagery. LiDAR point clouds provide sub-meter spatial accuracy for asset geolocation.
PII Redaction
Automated PII detection and redaction across all imagery with 95%+ accuracy — faces, license plates, and identifying information removed before storage.
Data Platform
Results stored in Snowflake with geospatial indexing. API layer serves detection results to city planning and maintenance systems.
Tech Stack
Compute
AWS ECS Fargate (48 parallel workers)
Data Warehouse
Snowflake
Language
Python
AI
Vision AI
Processing
LiDAR point cloud analysis
Analysis
Geospatial detection & mapping
Results
Asset types detected
PII redaction accuracy
Lines of production code
Automated tests
Parallel ECS workers
Full LA survey coverage