ML DevOps
Dynamic Risk Assessment System
An automated ML system that predicts client attrition risk for a company managing 10,000 corporate clients, with continuous monitoring, retraining, and API-based predictions.
The Challenge
A company managing 10,000 corporate clients needed a way to proactively identify accounts at risk of attrition. Manual monitoring was not scalable, and static models quickly became outdated as client behavior shifted over time. The goal was to build a dynamic system that could automatically detect data drift, retrain models, and serve predictions through a live API.
Our Approach
The system implements a complete ML lifecycle: automated data ingestion that detects and compiles new training data, model training with scikit-learn for attrition prediction, and a Flask-based REST API that serves predictions and performance metrics. DVC-style diagnostics generate dataset statistics and timing analyses to monitor model health.
Automation & Monitoring
Full process orchestration is achieved through cron job scheduling. The system automatically checks for new data, retrains models when data drift is detected, generates confusion matrix visualizations, and produces automated reporting documents. API endpoints expose model predictions, scoring metrics, and diagnostic summaries.
Results & Impact
The architecture delivers a self-maintaining risk assessment pipeline. New data is automatically ingested, models are retrained without manual intervention, and predictions are served with sub-second latency. The reporting layer ensures full visibility into model performance over time, enabling proactive client retention strategies.