Marketing Intelligence Platform
Four production ML models serving real-time predictions through a FastAPI backend, with automated monitoring and a feedback loop. Built end-to-end: data pipelines, model training, REST API serving, and operational dashboards.
System Architecture
Data Sources
UCI datasets
Bank Marketing + Online Retail II
Training Pipeline
scikit-learn / Prophet
Feature engineering + model fitting
Model Artifacts
Serialized models
joblib + pickle persistence
Serving Layer
FastAPI
REST endpoints with validation
Infrastructure
Hetzner VPS + Caddy
Systemd + auto-SSL reverse proxy
Monitoring
SQLite store
Prediction logging + feedback loop
Production Models
Predicts whether a customer will respond positively to a direct marketing campaign. The model ingests demographic, financial, and campaign-history features to output a calibrated conversion probability with SHAP-based feature attribution.
Clusters customers into behavioral segments using Recency, Frequency, and Monetary value analysis. Each new customer is assigned to the nearest cluster centroid with a confidence distance metric.
Generates multi-period revenue forecasts with confidence intervals using Facebook Prophet. Captures weekly and yearly seasonality patterns from historical transaction data to project future revenue trajectories.
Recommends products using Alternating Least Squares matrix factorization on purchase history. Supports both known-customer personalized recommendations and cold-start fallback to popularity-based ranking.