SpellML to SageMaker

The company

Whatnot is a social marketplace for collectibles and goods, where discovery is the product — what a buyer is shown decides what gets sold.

The problem

Their recommendation engine ran on SpellML, and the platform had stopped fitting the problem in three distinct ways: it struggled as data volumes and model complexity grew, its rigidity slowed experimentation, and there was no mature MLOps practice governing the model lifecycle.

The third one is the real constraint. A team can tolerate a slow platform; what it cannot tolerate is not knowing which model is in production or how it got there.

What I built

The migration centred on Amazon SageMaker:

• SageMaker Studio for development environments
• SageMaker Pipelines for orchestrated training workflows — data ingestion, feature engineering, training, evaluation as discrete reviewable stages
• SageMaker Endpoints for real-time model serving
• Terraform for the infrastructure, GitHub Actions for CI/CD

Models were deployed as real-time endpoints and load tested before they carried production traffic. The pipeline integrated with the MLOps tooling they already had rather than replacing it wholesale.

Result

Training and serving moved onto managed infrastructure that scales with the data instead of against it, and experimentation cycles shortened because standing up a variant stopped being an infrastructure task.

Cost optimisation came from the shape of the platform — pay-per-use pricing and spot instances for training — rather than from any one tuning exercise.

This engagement did not publish quantitative outcomes.