I am an analytics and insight leader with 7+ years of experience across customer strategy, marketing analytics, planning, budgeting, and P&L in large-scale ecommerce.
My work starts with the decision: who to retain, which intervention to use, where to invest, how much demand to prepare for, and how to measure incremental value. I use analytics, machine learning, causal inference, forecasting, and optimization as tools for answering those questions—not as endpoints.
|
Customer decisions Segmentation · Churn · Next purchase · Treatment selection |
Commercial decisions Incrementality · Budget allocation · Subscription economics |
|
Planning decisions Demand forecasting · Capacity planning · KPI review |
Delivery standard Reproducible pipelines · Baselines · Holdouts · Tests · CI |
Each public project uses synthetic data and reports results generated by committed code. Metrics validate the workflow under a controlled setup; they are not claims about production performance.
|
Decision: Who should receive which retention treatment when budget and channel capacity are limited? Executed evidence: The optimized policy produced 76.2% more simulated incremental net value than risk-only targeting on the untouched test period.
|
Decision: Where should the next unit of spend go when returns saturate and operating guardrails apply? Executed evidence: The optimized mix generated 10.2% more simulated incremental profit than the feasible historical mix, with 0.3% regret versus the simulation oracle.
|
|
Decision: Which behavioral and value profiles are stable enough to inform testable customer strategies? Executed evidence: Holdout ARI reached 0.971 versus 0.538 for an RFM baseline; mean bootstrap stability was 0.998.
|
Decision: Who is likely to miss the next purchase window implied by their own historical cadence? Executed evidence: Time-holdout PR-AUC was 0.602, with 1.60× lift and 32.0% recall in the top 20%.
|
|
Decision: How many orders should operations prepare for, and how should uncertainty change the capacity plan? Executed evidence: Untouched 56-day holdout WAPE was 4.9%; 80% empirical intervals achieved 80.8% coverage.
|
Decision: Did a campaign cause additional behavior—and did the gain exceed the full campaign cost? Executed evidence: CUPED estimated 7.0% order lift, while full cost reconciliation showed −0.8% incremental ROI and supported a do-not-scale decision.
|
- Define the business decision, action, and constraint before choosing a model.
- Compare against an understandable baseline and protect time-based or untouched holdouts.
- Separate prediction, causal impact, and business value so the claim matches the evidence.
- Build reproducible public case studies with synthetic data, automated tests, CI, and explicit limitations.
Next Purchase Recommendation · Agentic Business Review · Community Detection · Subscription Value & P&L