The ML project lifecycle is the end-to-end, business-facing view of a machine learning project — bigger than the technical workflow, because it starts before any data is touched and continues long after a model is deployed.
Lifecycle vs Workflow — What's the Difference?
The ML workflow covers the technical steps (data → features → model → evaluation). The project lifecycle wraps around that with the business and organizational stages that make a project succeed or fail.
The Five Lifecycle Stages
| Stage | Key Questions |
|---|---|
| 1. Scoping | What business problem are we solving? Is ML even the right tool? What does success look like, in business terms? |
| 2. Data | Do we have (or can we get) data that's relevant, sufficient, and reasonably representative? |
| 3. Modeling | The technical ML workflow — training, evaluating and tuning candidate models. |
| 4. Deployment | How will predictions reach real users — an API, a batch job, an embedded feature? |
| 5. Monitoring & Maintenance | Is the model still accurate as real-world data changes? When do we retrain? |
Why "Scoping" Is Usually the Most Skipped, Most Costly Stage
Teams often jump straight to modeling, and only later discover the business problem didn't actually need a predictive model (a simple rule or a dashboard would have worked), or that the model's output doesn't fit into how the business will actually use it. Time spent scoping — defining the target variable, the acceptable error rate, and how a prediction will be used — pays for itself many times over.
Practical Use Cases
This lifecycle applies whether the project is a churn-prediction model, a recommendation system, or a fraud detector — the stages are the same; only the details of the data, model and deployment method change. See ML Projects for full worked examples.
Common Mistakes
- Treating "deployment" as the finish line — the real cost of an ML system is often in stage 5 (monitoring and maintenance), not stage 3 (modeling).
- Starting stage 3 (modeling) before stage 1 (scoping) is actually finished — leads to technically good models that don't solve the real problem.
Interview Relevance
Q: "What happens after a model is deployed?" A candidate who only says "nothing, it's done" is missing the point — deployed models need monitoring for performance degradation and data drift, and a plan for when/how to retrain.
Practice Question
A team builds a highly accurate churn-prediction model, but six months later no one on the business side is actually using its predictions. Which lifecycle stage was most likely skipped or done poorly?
