AI lifecycle and data science process

The data science or AI lifecycle is a structured sequence of stages that turns a business problem into a working, monitored solution. A common framework is CRISP-DM.

1. Business or problem understanding: define the objective and the success metric.
2. Data collection and understanding: gather and explore data.
3. Data preparation: clean, transform and engineer features. This is usually the largest share of project time because real data is messy.
4. Modelling: train candidate models.
5. Evaluation: measure performance on data the model has not seen.
6. Deployment: release the model.
7. Monitoring and maintenance: watch for drift and retrain.

The lifecycle is iterative: results from later stages often send the team back, for instance to collect more data, fix a faulty join, or redefine the problem. Iteration happens inside one project.

Worked example: a model scores 98 percent on training data but 71 percent on held-out data. That gap signals overfitting, so the team returns to modelling to regularise or simplify rather than deploying.

A model whose accuracy falls months after launch because customer behaviour changed shows data drift; the fix is monitoring and retraining on recent data.

Exam trap: without a clear objective and success metric the team cannot judge whether the model is useful. Training score alone never justifies deployment.