The problem
A major construction group operating across Gulf states had recurring HSE violations — workers on site without helmets, high-vis vests, or safety gloves. Fixed cameras were expensive to install and maintain at temporary sites. Manual inspection by supervisors was inconsistent and slow.
The HSE director needed a tool any supervisor could use — pull out a phone, scan an area, and instantly know if anyone is non-compliant. No additional hardware. No cloud latency. No wait.
Why on-device was non-negotiable
Construction sites have poor connectivity
Large concrete structures, basements, and remote sites create dead zones. Any cloud-dependent solution would fail precisely where safety is most critical.
Privacy requirements
The client had GDPR and local privacy obligations. Worker images could not be transmitted to a cloud server without explicit consent infrastructure. On-device inference meant images never left the device.
Speed = usability
If the detection wasn't real-time, supervisors would stop using it. 30fps detection on a live camera feed was the minimum viable experience.
The solution
1
Custom YOLOv8 model training
Trained YOLOv8n (nano variant) on a 14,000-image dataset of construction workers with and without PPE across varied lighting, distances, and site types. Classes: helmet, no-helmet, vest, no-vest, gloves, no-gloves, safety-boots. Achieved 96.2% mAP@0.5 after 180 training epochs.
2
Dual export — TFLite + CoreML
Exported to TFLite (INT8 quantized, 6.1MB) for Android and CoreML for iOS. On Android mid-range devices the model runs at 22fps. On iPhone 13+ using the Neural Engine it hits 60fps — well above the 30fps requirement.
3
React Native app with live camera inference
Built a React Native app with a custom native module bridging to the TFLite/CoreML runtime. The camera feed is processed frame-by-frame. Bounding boxes and violation labels are drawn as an AR overlay in real time. Violations trigger haptic feedback and an audible alert.
4
Violation logging & reporting
Each confirmed violation is logged locally with a screenshot, timestamp, GPS coordinates, and assigned site code. Reports sync to the Django backend when connectivity is available. HSE managers see a live dashboard of violations by site, time, and violation type.
We deployed this on 40 sites in three countries. Violation rates dropped 61% in the first month. The supervisors actually use it — because it's fast and it works without signal.
— HSE Director, Construction Group, UAE
