Causal ML Remarketing Engine

The problem

High-volume remarketing can optimize clicks while still sending messages to users who would have purchased anyway. The product needed to decide who should receive a message, when it should arrive, and what to recommend while measuring true incremental impact rather than correlation.

My role

I built and led the ML remarketing engine for push and email across eight Latin American countries. The work combined causal modeling, personalization, serving architecture, and a sustained program of controlled experimentation.

Technical approach

• Causal uplift models with LightGBM and scikit-uplift to target persuadable users.
• Open-propensity audience filtering with XGBoost.
• Send-time models to optimize the moment of delivery.
• Price-aware recommendation ranking for message content.
• Recurring campaign strategies for remarketing, cross-sell, and buy-again use cases.

Experiment design

Together with partner teams, I co-designed more than 38 factorial 2×2 A/B experiments. Persistent holdout groups made it possible to distinguish incremental GMV from purchases that would have occurred without the intervention.

The program also treated negative evidence as useful: underperforming triggers were retired instead of being kept alive by vanity metrics.

Outcome

The ML optimizations drove approximately US$1 million per month in statistically significant incremental GMV versus holdout. Audience filtering improved conversion by 25%, while price-aware ranking improved recommended-item conversion by 19% versus control.

Results are rounded and described only at the level already disclosed publicly. No customer-level data or proprietary implementation detail is included.

What this work reinforced

For intervention systems, prediction accuracy is not the business objective. The key question is whether the decision changes an outcome that would not otherwise have happened.