A manufacturer has IP cameras at the end of each assembly line and a SageMaker-trained model to detect defects from still images. The factory must provide local feedback to workers when a defect is detected, even if internet connectivity is lost. A local Linux server hosts an API that delivers feedback. How should the company deploy the ML model so detection and local feedback work when offline?

The correct answer is to install AWS IoT Greengrass on the local server. Greengrass allows you to run AWS Lambda functions, Docker containers, and other components directly on edge devices, even when offline. This enables local inference using the SageMaker-trained model and direct communication with the local API for immediate feedback, satisfying the offline requirement. The first incorrect option, streaming to Kinesis Video Streams and using a SageMaker endpoint in AWS, relies on continuous internet connectivity, which violates the "even if internet connectivity is lost" requirement. The third incorrect option, using an AWS Snowball device, is designed for large-scale data transfer and edge computing in disconnected environments but is overkill and less agile for continuous, real-time local inference and feedback compared to Greengrass. The fourth incorrect option, Amazon Monitron, is designed for predictive maintenance of industrial equipment, not for general-purpose image defect detection from IP cameras, and it also relies on cloud connectivity for alarms and Lambda invocation.