Set up an Amazon Kinesis video stream from each IP camera to AWS. Use Amazon EC2 instances to take still images of the streams. Upload the images to an Amazon S3 bucket. Deploy a SageMaker endpoint with the ML model. Invoke an AWS Lambda function to call the inference endpoint when new images are uploaded. Configure the Lambda function to call the local API when a defect is detected.

Deploy AWS IoT Greengrass on the local server. Deploy the ML model to the Greengrass server. Create a Greengrass component to take still images from the cameras and run inference. Configure the component to call the local API when a defect is detected.

Order an AWS Snowball device. Deploy a SageMaker endpoint the ML model and an Amazon EC2 instance on the Snowball device. Take still images from the cameras. Run inference from the EC2 instance. Configure the instance to call the local API when a defect is detected.

Deploy Amazon Monitron devices on each IP camera. Deploy an Amazon Monitron Gateway on premises. Deploy the ML model to the Amazon Monitron devices. Use Amazon Monitron health state alarms to call the local API from an AWS Lambda function when a defect is detected.