BizSentinel
End-to-end ML platform for e-commerce customer intelligence
The Problem
Small and medium e-commerce businesses collect transactional data but lack dedicated data science teams to extract actionable insights from it.
The Solution
Three interconnected ML modules: Isolation Forest for anomaly detection, K-Means for customer segmentation, and LightGBM for churn scoring - each module feeding the next.
Impact
Gives SMB e-commerce operators early warning on churn, fraud-like behavior and customer segments so they can intervene before revenue is lost - across 97,896 customers, with privacy-preserving pseudonymization and differential privacy on the churn model.
Engineering Challenges
Module dependency chain
Anomaly scores and segment labels from unsupervised modules needed to feed as features into the supervised churn model without data leakage.
Privacy-preserving ML
Implementing differential privacy via diffprivlib while maintaining acceptable model performance required careful epsilon tuning.
Lessons Learned
Unsupervised outputs as supervised features
Feeding IsolationForest scores and KMeans labels into LightGBM meaningfully improved churn prediction - the pipeline design justified itself in the metrics.
SQLite is enough
Choosing SQLite over PostgreSQL eliminated operational overhead without any real cost at this data scale.