TechnologyExperimental AI report

GS Caltex Cuts Model Deployment Time to Hours with Amazon SageMaker

South Korean energy firm GS Caltex has slashed model rollout from a week to a few hours by pairing no‑code SageMaker Canvas models with SageMaker Pipelines…

Key points

  • GS Caltex split model work between process and ML engineers to speed deployment
  • SageMaker Canvas lets process engineers build models without code
  • SageMaker Pipelines automates deployment and schedules 30‑minute batch inference
  • Batch inference cuts inference costs compared with real‑time endpoints

GS Caltex, a South Korean energy and chemical company, announced that it now deploys machine‑learning models in a matter of hours rather than weeks by using Amazon SageMaker Canvas and Amazon SageMaker Pipelines.

The company separates model work between process engineers, who build and validate models in SageMaker Canvas without writing code, and ML engineers, who configure SageMaker Pipelines to deploy those models and schedule inference jobs.

No‑code model creation

Process engineers use SageMaker Canvas to create models that reflect their domain knowledge. The models are validated within Canvas before being exported to SageMaker Pipelines, preserving the data‑preprocessing flow that was built in the no‑code environment.

Automated deployment and scheduling

ML engineers register the exported models in the SageMaker Model Registry, approve them, and then deploy them as SageMaker model resources for Batch Transform. A five‑step inference pipeline—input preparation, preprocessing, target removal, batch inference, and result assembly—is defined in SageMaker Pipelines. Amazon EventBridge triggers the pipeline every 30 minutes using a cron expression, producing 48 scheduled runs per day.

GS Caltex reports that this batch‑based approach has lowered inference costs compared with continuously running inference endpoints, while also reducing the time required to move a model from development to production.

The reference implementation demonstrates the workflow with a sample loan‑classification model; real models would require configuration adjustments.