How do you run ML inference at the edge with AWS?

Train in the cloud, compile for the target hardware, and deploy models as versioned components via AWS IoT Greengrass, with staged rollouts, local buffering for intermittent connectivity, and telemetry flowing back for monitoring and retraining.

Edge inference is chosen for one of four reasons: latency that a network round trip cannot meet, bandwidth cost of shipping raw sensor data, operation during disconnection, or data-residency rules preventing raw data leaving the site. If none apply, cloud inference is simpler and usually better.

Architecture…