Stand up model serving on AWS for a computer vision inspection startup

Sightline Robotics does automated defect detection for packaging lines. Our two vision models currently run on a single beefy EC2 box that an intern set up, and it falls over whenever a second factory comes online. We need a proper serving setup: containerized inference with Docker, autoscaling on EKS, GPU node groups, blue-green model rollouts, and latency monitoring with alerts. You would own the infrastructure end to end - Terraform or CDK (your call, argue for it), CI that builds and pushes model images, a canary process for new model versions, and runbooks our two ML engineers can actually follow at 2am. Target: p95 inference under 200ms at 40 requests/second per site, and a new model version deployable in under 15 minutes without dropping requests. We estimate 15-20 hours a week alongside our team. Timezone overlap with US Central for at least 3 hours daily is required because rollouts happen during factory maintenance windows.