Looking at a dataset demonstrating carrier churn, we decided to create an algorithm to determine the probability a customer might leave T-Mobile. Factors include length of account, number of lines, usage, and payments left on devices.

What it does

Ideally, when a user opens the T-Mobile app, they will have a customized screen laid out according to what the Un-Churner algorithm predicts about the health of their account. If it appears the customer might be at risk for leaving T-Mobile, a non-invasive widget will appear at the top of the screen demonstrating a different phone plan or device, based on what appears to be the better solution to keeping them a happy customer with T-Mobile.

We were not able to fully implement an end-to-end demo of a customized front end, but we can demonstrate the algorithm in action.

How We built it

Data was used to train an algorithm that was built in Python. Flask runs our backend to work with Python. We wanted to use React, but that proved too ambitious in this limited time frame.

Challenges We Ran Into

Creating data, figuring what data we wanted. merging the algorithm to the front end. Initially, we were also very ambitious with the scope of the project. We wanted to not only calculate churn probability, but also based on different milestones in a customer's account history, provide tailored recommendations of plans, phones, and accessories. This would have been calculated by compiling sales history along with publicly available Amazon review data to see what other customers with similar purchase history are buying as well.

Accomplishments that We're proud of

Five developers with different skill sets coming together to learn new skills, new technology, and create a functional proof of concept in less than 24 hours.

What I learned

We learned a lot about how to work together with different experience levels and different skill sets. Planning what stack to use, how to make the data, and other engineering challenges threw wrenches in our plans over the day, but overall we learned how to overcome them in order to create our product.

What's next for The Un-Churner

Continuing to refine the algorithm to be useful not only to customers, but to T-Mobile customer care representatives. Eventually, it would be exciting to see more data to hone in on the accuracy of the algorithm such as demographic, location, account history, store visits, calls to care centers, etc. This could be bundled into a customer "grade" or "health" number on a scale that can show representatives easily how satisfied the customer might be. Further, this could be aggregated to a customer profile that can be pulled up by a representative, either in a call center or in store, to display key customer information such as plan, lines, number of children, the "health" factor, and any recent calls or visits to the store.

Built With

Share this project: