Self-Hosted MLOps on Amazon EKS
A closed loop on open-source tools in one private EKS cluster: data read in place, GPUs for training only, a fail-closed gate and a person's signature before any release.
A reference self-hosted MLOps platform, shown for a computer-vision workload. Images stay on the site's file storage and are read in place. Kafka and Airflow run ingest, training, test, deploy and drift workers; PyTorch trains on a GPU node group that exists for training only; MLflow tracks every run and model. A candidate passes a fail-closed evaluation gate and a person's approval before Triton serves it. Only basic AWS services are used, so the same design moves to another Kubernetes.
Point at a component to light its links; select it for what it does here and the practice that makes it right on AWS.
When to use it
- Data must stay where it is, and managed ML services are not allowed
- Domain experts who are not ML specialists must retrain models on new data
- Every verdict must be traceable to the model, data and labels behind it
How NGT delivers it
Built from code, reviewed at a gate by an NGT architect, and handed over with runbooks your team can run.
MLOps
