Kaya AI: Smart Glasses Safety Command Center

Inspiration

The construction industry is one of the most hazardous sectors in the world, yet on-site safety monitoring relies heavily on manual supervision, which can't be everywhere at once. We were inspired by the emergence of lightweight smart glasses and wanted to see if we could turn them into proactive, intelligent safety assistants. The challenge? Smart glasses have extreme thermal and battery constraints, making heavy on-device AI impossible. We set out to solve this. What it does Kaya AI acts as an intelligent "brain" for a worker's smart glasses. It provides a real-time hazard and anomaly detection pipeline. The system processes a live POV video feed, detects structural deviations or missing PPE (like hardhats or safety harnesses), cross-references them against safety protocols, and instantly alerts the worker and the centralized command center.

How we built it To overcome the hardware limitations of smart glasses, we engineered a "Split-Brain" architecture:

The Edge (Glasses): A lightweight, low-FPS ingestion script (OpenCV) running on the edge that constantly scans for movement or basic shapes, minimizing battery and thermal load. The Cloud/Server: A heavy FastAPI server powered by a Vision-Language Model (VLM) pipeline. When the edge detects a potential anomaly, it triggers the heavy cloud engine to process that exact frame and generate a contextual alert. The Command Center: A real-time React (Vite) dashboard simulating an industrial safety control room, connected natively via WebSockets to stream live alerts to site managers. Everything is containerized and orchestrated seamlessly using Docker. Challenges we ran into

Latency vs. Accuracy: Balancing the speed of the edge triggers with the accuracy of the heavy cloud model was tricky. We implemented strict frame-rate capping and synthetic data fallbacks to ensure the stream wouldn't crash. Container Networking: Ensuring seamless, real-time WebSocket communication between the edge simulator, the FastAPI backend, and the React frontend across Docker containers required careful network configuration and CORS middleware handling. Accomplishments that we're proud of

Successfully decoupling the "trigger" (edge) from the "analysis" (cloud) to respect the physical constraints of smart glasses. Building a highly responsive, real-time dashboard that instantly populates with critical and medium severity alerts without needing a page refresh. Orchestrating a multi-container GPU-accelerated environment that can be spun up anywhere with a single docker-compose command. What we learned We gained deep insights into edge-to-cloud architectures, optimizing video streams for low-bandwidth environments, and utilizing WebSockets for instant, bidirectional telemetry in industrial settings.

What's next for Kaya AI

Integrating a real-time Knowledge Graph (RAG) to cross-reference anomalies with specific site safety blueprints. Expanding the anomaly detection to include thermal anomalies and auditory hazards. Deploying the edge simulator onto physical Meta Ray-Bans or similar AR hardware for live field testing.