Before and after
Managing by hand and with MLOps Central Platform, compared
The jobs compared here: training and deploying an AI model with field data.
| Step | Before (by hand) | With MLOps Central Platform |
|---|---|---|
| Gather data | Download a zip from the labelling tool and copy it to a shared folder | Press Sync on an external source in Datasets (데이터셋). Only changed files are fetched |
| Train | Log in to the GPU server and type script arguments from memory | Resubmit the previous run with the same settings from the wizard. Only GPUs wired to the tenant are used |
| Compare results | Open log files and copy numbers into a spreadsheet | Select runs in Experiments (실험) and Compare — metrics, parameters and curves side by side |
| Keep the model | Send model_final_v3_real.pt over chat | Register it as a version in Model Registry (모델 레지스트리), linked to its run and dataset |
| Sign off | A verbal OK | An Approvals (승인) request → the owner approves or rejects on screen, and the decision is recorded |
| Roll out | Log in to every device and swap files | Record it in Deployments (배포) and send the Pull model command to the target devices from Edge Fleet (Edge Fleet). If something goes wrong, record a rollback and have the devices pull the previous version |
| Watch | Look into it when complaints arrive | Drift (드리프트) and Alerts (경보) tell you first, and the retraining wizard opens from that screen |
On screen
Each training run is recorded automatically in the experiment list with its headline metric.
A responsible person approves before anything reaches production.
The state of every field device is shown on one screen.
Drift metrics flag models whose input data distribution has shifted.