User guide
How to build and run models in the workspace, from datasets to field deployment
This chapter is a guide for people who work in the workspace. It covers one full cycle: upload data, train a model, compare the results and pick a good model, serve it after approval, send it to field devices (edge), and retrain when performance drops in operation.
For installing the platform, see Installation; for admin screens such as members, tokens and training runtimes, see Administration; for external training containers and direct MLflow integration, see Integration.
How to read this chapter
| What | When to read it | Where |
|---|---|---|
| Tutorial — ship your first model to the field | Once, the first time, following it in order (about 30 minutes) | Tutorial |
| Screen reference | When you want to know what a button or field on a screen does | The per-screen pages of this chapter |
| How-to | When you want to finish one common task quickly | How-to |
Workspace menu
After you sign in, the workspace menu appears in the left sidebar. Features turned off for the tenant and menus your role cannot use are hidden.
| Menu | What it does | Place in the flow |
|---|---|---|
| Dashboard (대시보드) | Summary of key metrics (KPI), recent alerts, model status, GPU slots and Edge Fleet | Overall |
| Experiments (실험) | MLflow experiments and run lists, run comparison, registering a run as a model | ③ Experiments |
| Training (학습) | New training (11-step wizard), training run list and detail, pretrained weights | ② Training |
| Model Registry (모델 레지스트리) | Model versions and stages (Staging, Production, Archived), production promotion requests, serving image builds | ④ Registry |
| Container Registry (컨테이너 레지스트리) | Browse and import images in the platform's built-in registry | Supporting |
| Pipelines (파이프라인) | Pipeline run history | Supporting |
| Deployments (배포) | Send model versions to edge devices, roll back | ⑦ Deployment |
| API Services (API 서비스) | Create serving endpoints, test them in the inference console | ⑥ Serving |
| Monitoring (모니터링) | Operational KPIs and trends | ⑨ Monitoring |
| Alerts (경보) | Handle active alerts; rules, channels and silences | ⑨ Monitoring |
| Approvals (승인) | Approve or reject production promotion requests | ⑤ Approval |
| Datasets (데이터셋) | Create datasets, upload files, sync external sources | ① Datasets |
| Edge Fleet (Edge Fleet) | Field device status, commands and policies | ⑧ Edge |
| Drift (드리프트) | Trends in input data distribution shift (PSI, KS), retraining | ⑩ Drift → retraining |
At the top right are the Tenant selector (테넌트 선택) (for example DEMO), the notification bell and the account menu. If you have admin permission, Admin console (관리자 콘솔) also appears at the bottom of the sidebar.
The whole flow
A model goes around the cycle below. When drift is detected, you go back to ② Training.
① Dataset ──▶ ② Training ──▶ ③ Experiments (compare runs) ──▶ ④ Model registry (Staging)
▲ │ production promotion request
│ ▼
⑩ Drift ─┘ ⑤ Approval ──▶ Production
▲ "Retrain this model" │
│ ▼
⑨ Monitoring/alerts ◀── ⑧ Edge devices ◀── ⑦ Deployment ◀── ⑥ Serving image build · API service| Step | Menu | Role required | Result |
|---|---|---|---|
| ① Create a dataset, upload files | Datasets | Developer or higher | A FROZEN dataset and its file list |
| ② Training | Training → New training | Developer or higher | One training run = one MLflow run |
| ③ Compare runs | Experiments | Anyone (view) | The chosen run |
| ④ Register the model | Experiments → Run → Register model | Developer or higher | A model version (Staging) |
| ⑤ Production promotion request → approval | Model Registry → Approvals | Approver, admin | A Production version |
| ⑥ Serving image build, API service | Model Registry → API Services | Operator, admin | Container image, inference endpoint |
| ⑦–⑧ Deployment and edge devices | Deployments | Operator, admin | A model running on the device |
| ⑨–⑩ Monitoring, drift | Monitoring, Alerts, Drift | Anyone (view) | Retraining decision |
The tenant admin sets a role for each member. There are five: viewer, developer, operator, approver and admin. Buttons you lack permission for are hidden on screen, and the server rejects the request too.