AI Project Lifecycle and Delivery
Domain 2 tests delivery of AI products through the CPMAI lifecycle: an iterative, data-centric path aligned with CRISP-DM, with explicit gates rather than a linear software build.
Lifecycle rules
- Business Understanding defines the decision, users, value hypothesis, success metrics, constraints, and non-ML baseline.
- Data Understanding checks provenance, quality, representativeness, labeling, and legal basis before modeling.
- Data Preparation builds reproducible cleaning, labeling, and feature pipelines; leakage and sampling defects are fixed here.
- Model Development starts simple, tracks experiments, and documents assumptions and versioned data.
- Model Evaluation uses technical and business metrics, error analysis, fairness checks, and acceptance criteria agreed earlier.
- Deployment covers integration, monitoring, rollback, human oversight, and ownership after go-live.
- Iteration is normal: evaluation can send the team back to data preparation or even business understanding.
Worked example
A churn model has strong AUC but poor lift in a key segment. The team does not tune the model first. It reviews data understanding, finds recent customers are underrepresented, returns to data preparation, and renegotiates the segment-level success metric before redeployment.
Traps
- Starting with model selection.
- Skipping the non-AI baseline.
- Declaring success on a single aggregate metric.
- Treating deployment as the end.
- Using test data repeatedly until it leaks into tuning.