SmartSite Vision — Construction Site Safety AI

Client: Construction Safety Corp | Industry: Construction Technology | Timeline: 14 weeks

Key Result: 73% incident reduction

Real-time computer vision system monitoring construction sites 24/7. Custom YOLOv8 models on Raspberry Pi 5 detect PPE violations, hazardous zone breaches, and unsafe worker proximity with 0.38s alert latency.

The Challenge

Construction Safety Corp operated three active job sites in New York with a combined workforce of 340 workers across multiple shifts. Despite mandatory PPE policies and regular safety audits, they averaged 2.3 recordable safety incidents per month — well above the industry benchmark of 0.8. Manual safety supervision was inconsistent: supervisors could only physically monitor one zone at a time, violations were frequently missed during high-activity periods, and incident reports were reactive rather than preventive. Insurance premiums had risen 38% year-over-year, and OSHA citations carried compounding financial and reputational risk. The client needed a scalable, always-on system that could proactively catch violations the moment they occurred — not after the fact.

Our Custom Solution

We engineered SmartSite Vision as a distributed edge-AI safety platform. Custom YOLOv8 object detection models were fine-tuned on a proprietary dataset of 48,000 labeled construction site images covering 11 PPE item classes — hard hats, safety vests, steel-toed boots, gloves, eye protection, and fall harnesses. The models run natively on Raspberry Pi 5 units mounted behind each IP camera, eliminating cloud-inference latency entirely. DeepSORT tracking assigns persistent IDs to workers across frames so the system monitors cumulative exposure in hazardous zones — not just point-in-time violations. When a violation is detected, a FastAPI event pipeline triggers a sub-400ms Slack/SMS alert to the site supervisor with a timestamped annotated image. The Next.js command dashboard provides live multi-feed monitoring, historical incident heatmaps, worker compliance scores, and automated OSHA-format incident logs. All edge data syncs to a central PostgreSQL + Redis store for trend analytics and insurance reporting.

System Architecture & Technologies

  • 14 × Raspberry Pi 5 edge nodes running YOLOv8 inference at 18 FPS per camera
  • Custom multi-class PPE detection model (mAP@0.5 = 0.91) trained on 48K site images
  • DeepSORT multi-object tracking for persistent worker ID across camera frames
  • FastAPI event bus with Redis pub/sub for sub-400ms alert dispatch
  • WebSocket-powered live dashboard with multi-feed monitoring up to 14 simultaneous streams
  • PostgreSQL with time-series extensions for 90-day rolling incident analytics
  • Automated OSHA 300/301 report generation from incident event logs

Technologies: YOLOv8 • OpenCV • DeepSORT • Raspberry Pi 5 • FastAPI • Next.js • Redis • PostgreSQL • WebSockets

Measurable Results & Outcomes

  • 73%: Reduction in recordable safety incidents (2.3 → 0.6/month)
  • 0.38s: Median alert latency from violation detection to supervisor notification
  • 98.4%: PPE detection precision — virtually zero false positives on-site
  • $340K: Annual insurance premium savings from documented compliance improvement
  • 14: Edge cameras deployed across 3 active job sites, 24/7
  • 0: OSHA citations in 6 months post-deployment (vs. 3 in prior 6 months)

Client Testimonial

"In six months we went from three OSHA citations to zero. The system paid for itself in the first insurance renewal cycle alone."

— Marcus D., VP of Operations, Construction Safety Corp

Explore More Case Studies