Prism: Predictive Intelligence Platform

Detailed guidelines are present on Prism User Guide (available in the Resources section).

Before You Launch

Refer to Section 3 (Prerequisites) and Section 3.4 (Before You Launch Checklist) of the user guide for detailed instructions on each step below.

• Activate Amazon QuickSight in your AWS account. Note your QuickSight username and IAM service role name (default: aws-quicksight-service-role-v0).
• Complete the CDK bootstrap steps in your account before deployment. See Section 3.3.
• Verify quota checks for SageMaker and Bedrock to ensure required service limits are met in us-east-1, or request increases before deploying. See Section 3.2.
• Optional: To view SageMaker ML pipelines visually, create a SageMaker Domain. See Section 5.6.

Launch Parameters

• Training Dataset Size Tier: Select your data tier before choosing instance types. The stack will fail if instances are too small. See Section 4.1 for the full tier table.
• Pipeline Schedule: Controls how often the batch pipeline runs (default: 1 hour).
• Minimum AUC Threshold: The model quality gate - the model only deploys if it meets this score on the test set (default: 0.65, range: 0.50 - 0.80).
• VPC Options: Three modes are available - USE_DEFAULT (simplest, uses AWS default VPC), CREATE_NEW (recommended for production, set NAT Gateways to 0 to use VPC Endpoints for lower cost), and USE_EXISTING (requires your VPC ID, private subnet IDs, and route table IDs). See Section 4.2 for full details.
• QuickSight Service Role Name: Find it in IAM > Roles > search "quicksight".
• QuickSight Username/Email: Find it in QuickSight > profile icon > Manage QuickSight.
• Admin Email: Email address to receive important product/subscription-related notifications.

Monitoring the Deployment

The stack will show CREATE_IN_PROGRESS until the installer signals completion. To monitor in real time, find the instance ID in the CloudFormation Outputs tab (InstallerInstanceId), and run commands as specified in Section 4.1 step 6 of the User Guide to monitor progress.

Uploading Data

After deployment, upload your CSV files to s3://prism-raw-<account>-<region>/churn-data/. The platform automatically detects uploads, infers the schema using Amazon Bedrock, and starts the pipeline based on the configured schedule. If the schedule is 1 hour, the pipeline runs in an hour. You can also execute the state machine prism-processing-workflow manually in Step Functions to trigger immediate processing. Review Section 2 for the expected schema and required columns before uploading.

The platform is designed around an at-least two-upload workflow:

First upload - historical data for training. Upload your full historical dataset. The pipeline runs end-to-end and trains a churn prediction model. See Section 5.3 for full details.

• Schema inference: ~1 minute
• Glue ETL (size dependent): ~2-5 minutes for typical datasets (1 MB-2 GB)
• SageMaker training pipeline (size dependent): ~20-50 minutes - monitor progress in SageMaker Console > Pipelines
• Once Glue completes, your historical analytics QuickSight dashboard will be available within 1-5 minutes Wait for the SageMaker training pipeline to complete successfully before proceeding.

Second upload - inference data. Upload the data you want predictions on. It goes through the same ingestion process (schema inference > Glue ETL), and once Glue completes it automatically triggers a SageMaker Batch Transform job. When inference finishes, a second Glue job loads the predictions and your QuickSight prediction insights dashboard is updated. See Section 5.4 for full details.

• Ingestion (schema + Glue): ~2-5 minutes
• Batch Transform: varies by dataset size - monitor in SageMaker Console > Batch Transform jobs
• Dashboard update: ~1-5 minutes after uploading file and predictions sheet will be available ~1-5 minutes after inference (batch transform) completes