summary: End-to-end customer lifetime value prediction system serving monthly batch forecasts for loyalty-program budget allocation at a major Latin American retailer.
problem: Loyalty budgets were allocated by gut feel — no data-driven forecast of which customers would generate the most long-term value.
approach: Multi-container Kubernetes pipeline on Vertex AI: feature engineering in BigQuery, LightGBM model with Optuna hyperparameter tuning, automated retraining with MLflow tracking and drift detection via Cloud Monitoring.
built: Feature store in BigQuery, training pipeline with automated model selection, batch prediction service with Cloud Monitoring alerts, and a data-quality validation layer maintaining 99.9% uptime SLOs.
Customer Segmentation & Causal Impact
id: PRJ · 02 · 2020-2022
status: ● production · enterprise
summary: ML-powered customer segmentation combined with causal inference to measure and optimize marketing campaign ROI across loyalty programs.
problem: Marketing campaigns targeted broad demographics — no mechanism to identify behavioral segments or measure true causal impact of interventions.
approach: Dual clustering pipeline: K-means for broad behavioral segments, DBSCAN for density-based micro-segments. A/B test framework with Statsmodels and SciPy for significance testing, diff-in-diff causal inference to isolate campaign effects from seasonal trends.
built: Automated segmentation pipeline with weekly refresh, A/B testing framework with power analysis and early stopping, and causal impact measurement feeding marketing budget decisions.
Energy Anomaly Detection
id: PRJ · 03 · 2024-2025
status: ● delivered · contract
summary: Data-integrity and anomaly-detection pipeline for energy analytics, translating statistical findings into actionable risk mitigation for grid operators.
problem: Energy meter data contained systematic errors and anomalies that silently corrupted downstream analytics — no automated quality assurance.
approach: Multi-layer statistical validation: distribution-based outlier detection with Statsmodels, hypothesis testing for systematic bias detection, and custom validation rules for domain-specific energy consumption patterns.
built: Automated data-integrity pipeline with anomaly scoring, statistical validation reports with confidence intervals, and an alert system flagging quality issues before they reach production analytics.