AI Project Lifecycle and Capstone Project
What the exam tests
This domain tests project discipline: framing a problem so it is measurable and ethical, preparing a dataset without leakage, selecting and evaluating a model, and documenting and presenting the result for a non-technical audience.
Core concepts
- Lifecycle phases (CRISP-DM style): business/problem understanding → data understanding → data preparation → modelling → evaluation → deployment, with iteration back to earlier phases.
- Problem framing: state the decision the model supports, the unit of prediction (student, crop, transaction), the label definition, the success metric, and the constraints. A vague goal ('use AI in school') is not a project.
- Dataset preparation: collect, clean, engineer features, split into train/validation/test before any fitting, and document sources and permissions.
- Model building: start with a simple baseline, then compare alternatives using the chosen metric; tune on validation data only.
- Documentation and presentation: record data sources, assumptions, model choice, metrics, limitations, and ethical considerations; present the problem, approach, result, and recommendation without jargon.
Worked example
Capstone goal: predict which Class X students risk dropping out. Framing: unit = student; label = dropout within one term; features must be available at prediction time, such as attendance, assignment submission, and distance from school. Metric: recall, because missing an at-risk student is costlier than a false alarm. Reasoning: including 'final exam marks' as a feature is leakage - it is only known after the outcome and would not exist when the counsellor needs the prediction. The model would score brilliantly in testing and fail in use.
Traps that cost marks
- Starting with the model or a favourite algorithm before defining the problem and metric.
- Data leakage: using features unavailable at prediction time, or scaling before splitting.
- Reporting only the best score and omitting limitations.
- Presenting raw accuracy to a non-technical audience without context.
- Ignoring consent and permissions when collecting student data.