SageMaker AI: the machine shop

Build

• SageMaker Studio — the web IDE: notebooks, code editors, and launchers for every other feature. Where ML engineers live day to day.
• Data Wrangler — visual data preparation: import, profile, and transform datasets with prebuilt transforms, then export the flow as a processing job.
• Feature Store — one pipeline, two read paths: an online store for low-latency lookups at inference and an offline store for building training sets. Ends the train/serve feature-skew problem.
• JumpStart — a hub of pretrained models and solution templates you can deploy or fine-tune as a starting point.

Train

• Training jobs — the core move: SageMaker AI spins up managed, ephemeral compute, pulls a container image (built-in or yours from ECR), reads data from S3 (or EFS / FSx for Lustre), runs training, writes model.tar.gz back to S3, and tears everything down. You rent the training run, not the cluster.
• Automatic model tuning (AMT) — hyperparameter search as a managed service: define ranges and an objective metric, and AMT launches coordinated training jobs (Bayesian, random, Hyperband, or grid strategies) to converge on the best configuration.

Track

• Model Registry — versioned model packages grouped by use case, each carrying metadata and an approval status. CI/CD pipelines promote only what a human (or a gate) marks Approved.
• Managed MLflow — a fully managed MLflow tracking server for experiments: log runs, compare parameters and metrics, and register winners without operating MLflow yourself.

Serve

Four inference options, one decision drill:

• Real-time endpoints — persistent, auto-scaling, low-latency HTTPS serving. The default for interactive traffic.
• Serverless inference — no instances to manage; great for intermittent or spiky traffic that tolerates cold starts.
• Asynchronous inference — queue-backed endpoints for large payloads or long-running requests; responses land in S3.
• Batch transform — score an entire dataset in a job, no endpoint left running afterward.