What Geo-MLOps does
Which stage of the model lifecycle each menu handles
After signing in you see the workspace menu on the left. Its order closely follows the model lifecycle.
Feature map
| Stage | Menu | What it does |
|---|---|---|
| Data | Datasets (데이터셋) | Upload images, annotations, time series, tables and point clouds, or sync them from an external DataOps system |
| Training | Training (학습) | Pick the task, model, dataset, hyperparameters and runtime in an 11-step wizard and train on the GPU cluster |
| Tracking | Experiments (실험) | Every training run keeps its parameters, metric curves and artifacts. Select several runs to compare them |
| Versions | Model Registry (모델 레지스트리) | Manage model versions and stages (Staging, Production, Archived) and build serving images |
| Images | Container Registry (컨테이너 레지스트리) | Store training and serving container images. Images built elsewhere can be imported as files |
| Review | Approvals (승인) | Approve or reject production promotion requests |
| Delivery | API Services (API 서비스) · Deployments (배포) | Run a model as a REST API (try it right away in the inference console), or deploy it to field devices and roll back |
| Field | Edge Fleet (Edge Fleet) | Register devices, see health, collected data and inference results, send commands and policies |
| Watch | Monitoring (모니터링) · Alerts (경보) · Drift (드리프트) | Key performance indicators (KPIs), alert rules and channels, input distribution shift and retraining |
| Automation | Pipelines (파이프라인) | Run history and step graph (DAG) of training and deployment jobs |
Administration
Besides the regular user menu there are two consoles.
- Admin console (tenant administrators): members and roles, tokens, audit log, and the tenant's execution environment such as training runtimes and serving builders
- System console (global administrators): tenant creation, users, shared images, system health, build cache, weights mirror, Prometheus and Alertmanager, Kubernetes cluster, nodes and GPUs
These are covered in Administration.
Connecting from outside
| What | How |
|---|---|
| Your own training code | Build a container image that follows the agreed rules, upload it, and it can be picked in the wizard |
| An in-house DataOps or labelling system | Implement three APIs and datasets are synchronised |
| Experiment tracking from notebooks and scripts | Log with the standard MLflow client (issue a token) |
| Field devices | The geo-mlops-sdk agent handles collection, sync, model delivery and inference logging |
| Automation from other systems | REST API (OpenAPI documentation included) |