MLOps
Models ship like code: versioned data, gated releases and a person who signs.
At a glance
Models in production need the discipline of code: versioned data, gated releases, watched serving and a person who signs. NGT builds that loop on open-source tools, self-hosted on Kubernetes when data must stay where it is.
Ingest, label, train, gate, shadow, serve, watch for drift and retrain, with one write-once audit log.
Open-source tools on Kubernetes, GPUs for training only, images read where they already live.
A fail-closed gate, human approval and a signed run manifest. Nothing is promoted automatically.
One cluster, one closed loop
A reference self-hosted MLOps platform on Amazon EKS, shown for a computer-vision workload.
- Data stays in placeData is read where it already lives; datasets reach S3 through a VPC endpoint.
- GPUs for training onlyCPU nodes run the API, workers and serving; GPU nodes exist for training.
- Fail-closed, then signedA candidate must pass a fixed benchmark, then a person signs the release.
- Drift retrains, never releasesDrift starts a candidate run that trains and tests, but never deploys.
Point at a component to light its links; select it for what it does here and the practice that makes it right on AWS.
Problems we solve
Five problems that keep models out of production, each with the architecture we use. Select one.
Data scientists waiting on tickets
Data scientists wait for ML workspaces and deployment slots through tickets, and every model reaches production by a different route.
- ML workspaces requested from the developer portal
- Model deployment from a template, in one click
- The same approvals and cost tracking as other resources
Data scientists request workspaces and deploy models themselves, inside the same guardrails as everyone else.
Point at a component to light its links; select it for what it does here and the practice that makes it right.
Explore MLOps
5 pages, each a few minutes to read.
Frequently asked questions
What customers ask first.
When data must stay where it is, or the platform must move between clouds or into a data centre. Where both apply, an automated check can block managed AI services.
Otherwise SageMaker, Azure Machine Learning or Vertex AI are often the faster start, and we build on them too.
Yes. A web interface covers datasets, labels, runs and approvals, so domain experts can retrain on new data. Release still needs the gate and a sign-off.
It depends on the model and the hardware. We agree a latency target per decision up front and measure it on your hardware before go-live.
No. Drift can start a candidate run automatically, but that run only trains and tests. Release needs the gate, shadow results and a person's signature.
The reference build runs on Amazon EKS. The same design runs on AKS, GKE, your data centre or FPT AI Factory; we also build on SageMaker, Azure Machine Learning and Vertex AI.
Application and pipeline code, Terraform, dashboards, runbooks and a model card per model, in your repositories.
