Predictive Modeling for Automotive Retail

Use a centralized data lake account to accelerate development of new use cases.

Strip personally identifiable information. Aggregate data to obfuscate dealer specifics and prevent bias for or against OEM dealers.

Use AWS Database Migration Service to replicate on-premises databases that are not available through a data lake account.

Preprocess data with AWS Glue PySpark Transforms and output the results into a primary table for model training.

Push the training input primary table from the data ingestion pipeline at regular intervals and store it in Amazon Simple Storage Service.

Trigger the model training pipeline when the primary table changes. The pipeline includes hyperparameter tuning, validation, and fit. Step Functions controls this pipeline by using AWS Deep Learning Containers.

Trigger a container build on model source commits. Store the container in Amazon Elastic Container Registry.

Store versioned model outputs in Amazon S3, including the training report and evaluation.

Use Amazon Simple Notification Service to notify an administrator for review before triggering deployment with Step Functions.

Serve inference requests through Amazon API Gateway, AWS Lambda, and an SageMaker AI endpoint that uses the trained model container.

Build the inference dashboard by using AWS Amplify. Host static content on Amazon S3 and serve it through Amazon CloudFront.