Describe a project where you applied machine learning to solve a real-world problem. What was the problem, and how did your solution impact the business?

Illustrative answer

This is a fictional churn-prediction project. Its counts illustrate evaluation; they are not results from a real company.

Our retention team could review only a limited number of accounts and wanted a better way to prioritize that review. I helped build a churn-risk model using information available at the review date. I defined the later churn window with the business owner and excluded fields recorded only after cancellation.

We compared ranking by recent usage with a regularized logistic-regression model using usage and support-history features. We trained on earlier cohorts, used a separate validation period for choices and reserved a later cohort for the final check. Preprocessing was fitted on training data. I examined missing inputs and relevant customer groups as well as the overall score.

In the held-out cohort of 1,000 accounts, 200 later churned. Of the model's top 100 accounts, 60 later churned: precision at that capacity was 60%, and recall was 60 of 200, or 30%. Selecting 100 accounts at random would include 20 churners on average in this cohort, so the model selected three times that expected number. It still missed 140 of the eventual churners.

The business owner approved a bounded workflow evaluation with human review. We had evidence of useful ranking, not proof that contacting those accounts would prevent churn. Retention impact, contact costs and possible customer harm still needed evaluation, ideally through an appropriate controlled comparison. My contribution changed the prioritization decision while keeping prediction quality separate from an unmeasured revenue or retention claim.

Follow-up to practice

Why might the highest-risk account be different from the account most likely to benefit from an intervention?

Technical background: scikit-learn on decision-time information and leakage.