At Amazon, I adapted pretrained Transformers for Ads measurement, building a modeling pipeline from 10M temporally aligned searches, ad impressions, and purchases. The framework improved off-Amazon purchase prediction ROC-AUC by 7.5% over production baselines and extended to customer-level Shapley attribution.
At Wayfair, I enhanced the Multi-Touch Attribution system with an end-to-end feature enrichment pipeline. Evaluations validated against 6 geo-lift experiments reduced attribution error by $3.13M (10.7%), while productionized Airflow workflows reduced model runtime by 14%.
My research includes LLM evaluation, spatio-temporal demand forecasting, and reinforcement learning for bus operations. In a bus-holding project, the framework reduced headway variability by 32.0%, bus bunching by 69.2%, and passenger waiting time by 24.0% over the best calibrated rule-based baseline.
