The Challenge
FinBridge was losing high-value lending clients without warning. Their account managers had no early-signal system — they only found out a client was leaving after the fact. Retention campaigns were reactive and expensive. The data existed (transaction history, login frequency, support interactions) but no one had modelled it.
The Solution
I engineered a gradient-boosted churn model (XGBoost) trained on 24 months of client behaviour data. Feature engineering focused on recency-frequency-monetary signals plus rolling averages of product usage. The model was wrapped in a FastAPI endpoint deployed on Railway, returning a churn probability score for each client every 24 hours. Scores above 0.65 trigger an automated Slack alert to the assigned account manager.
Results
87%
Recall on true churn cases
0.81
AUC-ROC score on held-out test set
2 wks
From kickoff to production deployment
3×
Improvement in retention campaign ROI
Tech Stack
“We had the data sitting there for years. Charles turned it into a live warning system in two weeks flat. Our account managers now have the heads-up they need to actually save the relationship.”
James Frimpong
Chief Risk Officer, FinBridge Capital