MLOps Central Platform does not replace MLflow. It embeds MLflow and adds the layer that running models in production needs on top of it.

Why MLflow

Tracking experiments and managing models by hand causes these problems:

  • Reproducibility: environments differ between personal machines, so it is hard to trace which data and parameters a run used.
  • Version management: versions and scores are compared in spreadsheets and file names.
  • Deployment and monitoring: training and deployment are manual, and a drop in performance after release goes unnoticed.

The platform uses MLflow as the basis for experiment tracking and the model registry.

ReasonDetails
De facto industry standardOpen source that provides experiment tracking and a model registry together
Easy to adoptA standard Python API: a few lines added to training code
Framework-neutralLogs PyTorch, YOLO, scikit-learn and others the same way
Proven storageRuns on widely used stores such as PostgreSQL and S3

A setup with MLflow alone

In a standard MLflow setup, clients log experiments and register models through the tracking server.

MLflow service architecture: the Python client, CLI and web browser reach the MLflow Tracking Server over its REST API; metadata is kept in PostgreSQL and artifacts in S3-compatible storage (diagram labels in Korean)

  • Users: ML developers work through the Python client (log_params, log_metrics, log_artifact, register_model), the CLI or a web browser.
  • Service: a single MLflow Tracking Server handles experiment tracking, the model registry, the artifact service and the web UI.
  • Storage: metadata such as experiments, runs and model versions goes to PostgreSQL; large files such as model weights, images and logs go to S3-compatible storage.

The default setup has no authentication and no isolation between teams. Anyone who knows the address can see and change every experiment and model.

Where MLflow alone falls short

GapWhat goes wrong
SecurityWith no authentication, models and experiment data are exposed
GovernanceThere is no approval step for putting a model into production, and no audit trail of who did what
OperationsMLflow does not deploy, serve, monitor (detect drift) or retrain (CT)

The layer MLOps Central Platform adds

A platform layer in front of the MLflow core adds multi-tenancy, approval, deployment and observability.

MLOps Central Platform architecture: a platform layer (auth and RBAC, MLflow gate, approval workflow, audit log, deployment management, drift watch, monitoring) sits in front of the existing MLflow core and shares the execution, observability and storage services (diagram labels in Korean)

The red dashed box in the diagram is what was added. The MLflow Tracking Server is reachable only through the platform gate and is never exposed directly.

AddedWhat it doesMore
MLflow gateAuthenticates the caller with a tenant token and gives each tenant its own MLflow workspace. Developers keep using standard MLflow codeUse MLflow from a notebook
Role-based accessFive roles (VIEWER, DEVELOPER, OPERATOR, APPROVER, ADMIN), an approval workflow and an audit logRoles and permissions
Operations automationEdge deployment and rollback, monitoring, alerts, drift watch and retrainingWhat MLOps Central Platform does
One UIFrom experiments to deployment and monitoring, in one placeUser guide

Comparison

MLflow aloneMLOps Central Platform
Experiment tracking and model registry✅✅ (the bundled MLflow, unchanged)
Login and authenticationNoneLogin (JWT); MLflow uses tenant tokens
Team (tenant) isolationNoneA workspace and a Kubernetes namespace per tenant
PermissionsNoneFive roles, eleven permissions
Promotion to ProductionAnyone, immediatelyAlways through an approval request; safety-critical models need two approvals
Audit trailNoneWho did what is recorded and can be exported as CSV
Running trainingBy handSubmitted from the wizard and run as a Kubernetes Job on the cluster
Serving and edge deploymentBy handAPI services, deployment to field devices, rollback
MonitoringNoneMonitoring, alerts, drift (PSI, KS)

Written for the platform as of 2026-10-02.

© MLOps Central Platform