Automated Hazard Prediction Solution Reaches 98% Accuracy in Care Facilities

To execute this, our team strategically utilized the AWS architecture, where SageMaker orchestrated the model, with AWS Lambda serving as the event-driven compute layer and utilizing a suite of LLMs. For this solution, we selected the LLM approach for its superior simplicity and adaptability in differentiating image labels and classes, which was crucial for accurate hazard identification.

Our experts trained the LLM with four classes of sample images (clear vs. hazard-present) to teach the system, while utilizing Lambda and Bedrock for inference. In Bedrock, we obtained the output label classification for images and returned the results to Lambda.

Crucially, our rigorous validation process demonstrated high viability across various LLM configurations, confirming the technical feasibility of the solution. This result demonstrates the model’s high adaptability, as it can be easily retrained on diverse datasets—whether for fire detection or warehouse analytics—by simply providing a new image feed and an updated script to adjust labels and prompts.

As a result, the solution can be trained in various ways and provide businesses with a precise prediction model that detects unusual patterns, warns employees about potential identifiers, and prevents incidents before they occur.