MLOps · 01 of 5
The closed loop
Ingest, label, train, gate, shadow, serve, watch and retrain, with rules no one can skip.
IngestLabelTrainGateShadowServeMonitorRetrainClosed loop
01
A closed loop, not a pipeline
Eight stages on one platform. Every stage writes to the same audit log; the gate cannot be skipped.
Human-gated loopretraining is automatic; releasing is not
Step 1 of 8
Ingest
Data read where it lives; duplicates removed by hash.
Always onWrite-once event logAccess per project and per roleA model card for every model
02
Six rules in the core code
Written into the platform's code, so no operator, script or deadline can skip them.
03
What happens when the data drifts
Retraining starts by itself. Releasing does not: the loop stays open on purpose.
1 of 4
Drift detected
Incoming data no longer looks like the training data.
04
Ad-hoc ML versus the closed loop
Where many teams start: manual review cannot keep up, and no prediction can be traced.
Before Closed loop
Datasets✕Copies without versions.✓De-duplicated, versioned datasets and labels.
Retraining✕When someone finds the time.✓Started by drift; candidate runs train and test only.
Release✕Weights copied by hand.✓Fail-closed gate, sign-off and a signed run manifest.
Rollout✕Straight to production.✓Shadow first, one-click rollback.
Traceability✕Nobody knows which model made a call.✓Every verdict traces to its model, data and labels.
Rare defects✕Models wait for enough defect examples.✓Anomaly models learn what a good part looks like.
SourcesAI Modeling & MLOps case study, NGT (Sep 2026)AI-First DevOps Transformation: Services & Solutions, NGT (Oct 2026)NGT Service Offering v3.2 (Jul 2026)Machine Learning Platform with Kubeflow, NGT reference deckPlatform Engineering v1.2, NGT (Sep 2026)Data Platform Ops & DataOps Service Catalog, NGT (Sep 2026)FPT AI Factory service offering, FPT Smart Cloud (Nov 2025)