Ehsan Shafiee

- Built and launched Dynamic FAQ, an LLM-driven product Q&A system that generates product-specific FAQs from merchant catalogs, historical support logs, and site content, grounded in real user demand via clustering of historical support questions. - Designed and productionized a contextual Thompson-sampling bandit to personalize and optimize FAQ selection using user, session, and product context, driving ~10× FAQ engagement and ~40% organic revenue lift in pilot merchants. - Designed and ran switchback (time-randomized) experiments to measure causal lift under production constraints, enabling reliable attribution of engagement and revenue impact. - Designed a layered recommendation architecture separating content-based retrieval (cold start & coverage), collaborative filtering (preference signals), and online optimization via contextual bandits. - Led the ML core of Smart Re-Order, applying gradient-boosted survival modeling (Cox-style loss) to predict time-to-next purchase and trigger personalized reorder reminders, achieving C-index 0.78 and ~30% improvement in targeting accuracy, with adoption by ~30 paying merchants. - Helped shape company-wide ML direction while remaining fully hands-on in coding, modeling, and production implementation.

- Led a 7-person cross-functional team to build a hierarchical Bayesian regression model estimating clients’ total core cash (“wallet size”) across banks. The model inferred latent demand from partial observations using revenue–industry segmentation and firmographic and financial signals, supporting data-driven cash-management sales decisions.

- Built an agent-based Monte Carlo simulation using rolling 60-day EHR and staffing data to perform counterfactual and causal analysis of infusion center operations, simulating ~10K scenarios per site to identify drivers of patient-flow inefficiencies and end-of-day overruns. - Supported pilot deployments across infusion centers, where 9 of 12 sites reduced end-of-day overruns by ~30 minutes on average; the simulation capability was later productized as a self-service feature.

- Built and productionized a dynamic-programming–based labeling pipeline using PySpark to generate position-bias–corrected relevance labels, enabling scalable supervision for global e-commerce search ranking models. - Implemented a non-parametric sequential A/B testing framework to support always-valid inference under continuous monitoring, improving decision reliability and accelerating experimentation for complex, non-Gaussian metrics. - Owned profit as a guardrail metric for online experimentation, building large-scale data extraction pipelines and deriving empirically grounded revenue–profit tradeoffs to support business-safe experiment decisions. - Co-developed a hierarchical demand forecasting system for ~140K SKUs across ~4.6K U.S. stores; personally owned seasonality modeling, out-of-stock imputation, and cross-item/store pooling, contributing to double-digit MASE improvements and reduced stockouts in pilot markets.

Research focus: Reinforcement Learning, Stochastic Systems, Bayesian Decision-Making Dissertation: Bayesian Decision-Making with Information Choice -Modeling Bayesian statistical decision making with information choice in dynamic and sequential set-ups. The framework connects artificial intelligence, privacy constrained data analysis, network control and behavioral decision making in economics. - Analytical solution for models of Rational Inattention with extensions to filtering and estimation and partially observable Markov decision processes.

Dynamic stochastic modelling of electricity market in the presence of renewable energy

Theoretical and empirical modeling of investment demand in real-estate market

Ranked 3rd among near 500,000 participants in national university entrance exam to be accepted to the most prestigious college program in the country, Electrical Engineering at Sharif University of Technology.