• 

  cover-image

Inspiration

Workplace safety inspections are often manual, reactive, and difficult to scale. Existing AI solutions focus on detecting hazards but rarely explain why a situation violates safety regulations, while many require sending sensitive video to the cloud. We wanted to build a privacy-preserving, explainable, and trustworthy AI system that assists safety professionals instead of replacing them.

What it does

Serenix AI is a privacy-preserving Edge AI platform for workplace safety compliance.

It provides two complementary workflows:

• Web Inspection: Inspectors capture or upload images of workplace assets (fire exits, extinguishers, electrical panels, etc.), and the system evaluates compliance against OSHA/NFPA regulations.
• Continuous Edge Monitoring: NVIDIA Jetson Orin processes CCTV streams locally to detect PPE violations, restricted-area intrusions, falls, and other workplace hazards without transmitting raw video.

The platform combines computer vision, deterministic compliance reasoning, Qwen-powered AI, and Retrieval-Augmented Generation (RAG) to generate explainable compliance reports while keeping humans in the decision loop.

How we built it

Serenix AI combines edge computing, computer vision, and AI reasoning into a modular architecture.

Tech Stack

• Frontend: React + Vite
• Backend: FastAPI
• Edge AI: NVIDIA Jetson Orin + DeepStream
• Computer Vision: YOLOv11 + OpenCV
• AI Models: Qwen LLM & Vision Language Models
• RAG: LangChain + FAISS
• Database: PostgreSQL
• Messaging: MQTT
• Deployment: Docker on Alibaba Cloud

Instead of relying on an LLM to determine compliance, we use a deterministic rules engine to evaluate OSHA/NFPA standards. Qwen is responsible for generating explainable reports, retrieving relevant regulations, and answering compliance-related questions.

Challenges we ran into

• Designing a privacy-first architecture that keeps sensitive video on the edge.
• Combining deterministic compliance rules with AI-powered reasoning.
• Supporting unreliable network connectivity through offline edge processing.
• Integrating multiple AI components while maintaining explainability and auditability.
• Building a scalable architecture suitable for industrial deployment.

Accomplishments that we're proud of

• Built a modular privacy-preserving Edge AI architecture.
• Combined computer vision with deterministic compliance reasoning.
• Integrated Qwen-powered explainable AI using RAG.
• Designed a human-in-the-loop workflow for trustworthy decision-making.
• Developed a scalable foundation for smart workplace safety across industries.

What we learned

This project reinforced that trustworthy AI requires much more than accurate predictions. Privacy, transparency, deterministic decision-making, and human oversight are essential for safety-critical systems. We also gained valuable experience integrating edge computing, computer vision, LLMs, and cloud infrastructure into a production-oriented architecture.

What's next for Serenix AI

• Expand support for additional OSHA, NFPA, ISO, and international safety standards.
• Add pose estimation and activity recognition for advanced hazard detection.
• Scale deployment across multiple edge devices and industrial sites.
• Introduce federated learning for privacy-preserving model improvement.
• Enhance multilingual compliance reporting using Qwen.
• Develop predictive safety analytics and risk forecasting.
• Continue evolving Serenix AI as an open-source platform for trustworthy AI-powered workplace safety.

Built With

• yolo