MLOps platform for solar generation forecasting

Goal

Forecast daily generation for thousands of inverters more accurately than the physics estimate, to run the grid and trade energy with less uncertainty.

What I did

• Designed the Dagster orchestration: one asset per lakehouse stage, schedules and sensors, and full lineage from replication to inference.
• Built YAML pipeline configuration with overrides from the Dagster UI, so strategies and tier thresholds change without code changes.
• Integrated Dagster with MLflow to log every training run and promote the champion model.
• Designed the environments: the same code runs on local Docker Compose and on EKS, with Terraform, Helm, Karpenter on spot instances, separate IAM roles and encrypted S3.