AWS ships per-camera vision pipeline with SageMaker Neo

The post is an update to a 2023 series on MLOps for visual quality inspection at the edge, now ported to the SageMaker Python SDK v3 and replacing the since-deprecated SageMaker Edge Manager with SageMaker Neo for edge optimization. The sample deploys via the AWS CDK; the orchestration layer admits at most five pipeline executions at a time by default, which keeps concurrent SageMaker training jobs inside the standard account quota and means no service quota increase is needed.

A design choice worth noting is that every quality gate scores the exported ONNX artifact rather than the in-memory PyTorch model, so the artifact that is gated is the artifact that ships. The classifier branch uses balanced accuracy plus defect recall because roughly 90 percent of images are good, which makes plain accuracy misleading. The Model Registry keeps packages from all cameras in shared groups and recovers per-camera history by filtering on camera metadata, because ListModelPackages cannot filter on customer metadata server-side. On a camera's first run the fetch finds nothing and training falls back to cold-start initialization; later runs resume from that camera's own prior model.

The sample also documents two SDK v3 behaviors that are not listed as breaking changes. The SDK validates the pipeline execution role against live IAM at synthesis time, so a CDK app cannot create the role and the pipeline in one deployment; the sample's two-phase deploy handles this with a context flag. The role validator also simulates ECR pull permissions without a resource ARN, so a role scoped to a single ECR repository fails validation even though it would work at runtime. Compilation itself incurs no charge; you pay for the processing instance that drives it and the S3 storage of the outputs.