Implementing and Handing Off an AI Solution or Environment
Domain 3: Implementing and Handing Off an AI Solution or Environment
This domain tests whether you can move a model into production, operate it safely, and hand it off so others can run it. The exam focuses on deployment patterns, CI/CD and MLOps, monitoring, security, documentation, and governance. Implementation is not finished when the model scores well offline; it is finished when the service is observable, maintainable, and owned.
Implementation rules
- Choose a serving pattern: batch for scheduled scoring, real-time API for low-latency requests, streaming for continuous events, and edge for offline or bandwidth-constrained sites.
- Package the model with its preprocessing and runtime environment. Training-serving skew occurs when features are computed differently in production than in training.
- Use version control for code, data, and model artifacts; a model registry for approvals; and CI/CD/CT pipelines to test, deploy, and retrain.
- Monitor service health, input drift, prediction drift, and outcome quality. Define alert thresholds and a rollback path.
- Secure secrets, apply least-privilege access, encrypt data, and keep audit logs. Document model cards, API contracts, SLAs, runbooks, and ownership.
Worked example
A bank deploys a fraud model as a real-time API. The model must score transactions in under 150 ms, so the team precomputes customer aggregates and serves only the needed features. They containerize the model and preprocessing, register the image, and use a canary release with 5% traffic. Monitoring tracks latency, error rate, feature drift, and fraud capture rate. If drift exceeds threshold or precision falls, traffic rolls back. Handoff includes an API contract, model card, runbook, alert playbook, and named owner.
Traps
Deploying without monitoring; recomputing features differently in production; hardcoding credentials; skipping rollback; treating a notebook as a deployable artifact; and handing off without documentation or ownership. Also watch for ignoring regulatory logging and access controls.