ChurnSense AI

Inspiration

Businesses usually find out a customer is unhappy only after they cancel. Retention teams also don't know whom to contact first, so they waste time on low-value customers while high-value ones quietly leave. We wanted a tool that turns customer data into clear decisions: who is at risk, why, and what to do about it.

What it does

ChurnSense AI predicts which customers are likely to churn and turns that into action:

Predicts a churn probability for every customer
Shows the top reasons customers leave (churn drivers)
Estimates the revenue at risk for each customer
Recommends a specific retention action, such as a payment plan, a priority support call, a re-engagement email, or a yearly-plan discount
Provides a dashboard with filters and a downloadable CSV action list, so the team can start working right away

How we built it

Data: We generated realistic subscription-customer data (tenure, monthly charges, support tickets, late payments, logins, contract type). The app can also take an uploaded CSV.
Model: A Random Forest classifier predicts the probability of churn for each customer.
Insights: Feature importance shows which factors matter most.
Business impact: Revenue at risk for each customer is calculated as

$$\text{Revenue at Risk} = P(\text{churn}) \times \text{Monthly Charges} \times 12$$

Recommendations: Rule-based logic maps each at-risk customer's behavior to the best action.
Dashboard: Streamlit shows the metrics, charts, and a ranked list of customers to save first.

What we learned

A prediction alone isn't useful unless it leads to an action, so ranking by revenue at risk matters more than ranking by probability alone.
Feature importance makes the model easier to explain to non-technical teams.
Simple, explainable rules can work well alongside machine learning.
Deployment details such as requirements.txt matter as much as the model.

Challenges

Building a complete project (data, model, insights, and dashboard) within a few hours.
Choosing sample data realistic enough to show meaningful churn patterns.
Fixing a ModuleNotFoundError for scikit-learn during deployment, which came down to a missing or misplaced requirements.txt.
Turning model output into advice that a business user can understand and act on.

Results

On our sample data the model reached about 70% accuracy and an AUC of 0.74. Since the data is synthetic, these numbers show the pipeline works rather than real-world performance.