COMP 3212 · Machine Learning on AWS: Deployment and Monitoring

Led by Norbertian Cybernetics Simulacrum

SageMaker deployment, CI/CD with CodePipeline, Model Monitor, Clarify, and the full responsible AI governance stack. Based on AWS documentation and MLOps practice.

1. Module 1

   Deploying ML Models

   Led by Norbertian Cybernetics Simulacrum

   The question

   Real-time endpoint (always-on, low-latency), serverless (scales to zero, cold start), asynchronous (large payloads), batch transform (whole datasets). Fraud detection at 100ms, monthly batch scoring, a document API called 20 times per day — which mode for which?

   Outcome

   The student can describe four inference modes and select the appropriate one.

   Sub-units

   1. ○ 1.1 Choose a Deployment Mode Open this unit → ✓ ↻
   Begin the course → Print certificate of completion
2. Module 2

   Infrastructure as Code

   Led by Norbertian Cybernetics Simulacrum

   The question

   Manually clicking through the console is not reproducible. CloudFormation declares the desired state in YAML; CDK writes infrastructure in Python. Docker containers package training and inference code. Why does IaC prevent the staging/production divergence that causes production incidents?

   Outcome

   The student can describe IaC principles and the Docker container deployment pattern.

   Sub-units

   1. ○ 2.1 IaC for ML Open this unit → ✓ ↻
   Open this module → Print certificate of completion
3. Module 3

   CI/CD Pipelines for ML

   Led by Norbertian Cybernetics Simulacrum

   The question

   Code push → CodeBuild runs tests → CodeDeploy updates production. SageMaker Pipelines is the ML equivalent: a DAG of preprocessing, training, evaluation, and deployment steps. Design the DAG for a monthly churn model retrain — including the condition step that stops deployment if accuracy is too low.

   Outcome

   The student can describe CI/CD concepts and trace code through a SageMaker Pipelines DAG.

   Sub-units

   1. ○ 3.1 Design an ML Pipeline Open this unit → ✓ ↻
   Open this module → Print certificate of completion
4. Module 4

   Production Monitoring

   Led by Norbertian Cybernetics Simulacrum

   The question

   Data quality drift, model quality decline, bias drift, SHAP attribution drift — SageMaker Model Monitor tracks all four. CloudWatch tracks infrastructure. X-Ray traces requests. CloudTrail logs API calls. Design a monitoring plan for a fraud detection model: which check, at which threshold, with which response?

   Outcome

   The student can describe Model Monitor's four monitoring types and design a production monitoring plan.

   Sub-units

   1. ○ 4.1 Design a Monitoring Plan Open this unit → ✓ ↻
   Open this module → Print certificate of completion
5. Module 5

   Responsible AI and Cost Management

   Led by Norbertian Cybernetics Simulacrum

   The question

   A bank's loan model is rejected by a regulator for lacking explainability documentation. Write the model card: bias checks, monitoring plan, alert response, individual explanation method, and remaining legal risks. Spot instances are 70% cheaper than on-demand — what is the cost of that discount?

   Outcome

   The student can write governance documentation for a production ML model.

   Sub-units

   1. ○ 5.1 Final Essay: Production ML Responsibility Open this unit → ✓ ↻
   Open this module → Print certificate of completion