Code for the paper by Falet et al. (2022) "Estimating treatment effect for individuals with progressive multiple sclerosis using deep learning"

Developing and Evaluating Causal Inference Methods for Pragmatic Trials

This project focuses on Exploratory Data Analysis (EDA) and Uplift Modeling to predict customer response to marketing campaigns.

Undersmoothing Causal Estimators with Generative Trees

In this project, we will work on replicating the results of a paper by Imbens & Gelman (2019). The paper states three arguments why high-order polynomials should not be used in regression discontuinity design.

[ICML 2026] Gradient-Based Causal Tree Ensembles for Heterogeneous Treatment Effects

Provides a convenient wrapper function for data analysis with regression discontinuity design (especially discrete running variables) as an identification strategy.

This project potray's my capstone work. My work involved cleaning, manipulation and analyzing huge data sets and implementing ML models for predictive analysis. This analysis demonstrates that data-driven credit interventions can meaningfully reduce financial exposure while balancing customer retention.

End-to-end MLOps pipeline for online causal inference: DoWhy/EconML DML models trained on Databricks, served on Azure Kubernetes via CI/CD.

Production-grade causal uplift modeling on 14M rows, benchmarks S-Learner, T-Learner, and FT-Transformer challengers on the Criteo dataset, with Optuna tuning, MLflow tracking, FastAPI + Docker + Google Cloud Run serving, and a Streamlit dashboard.

Zero-to-hero notebooks on causal inference and experimentation for global LLM rollouts: synthetic control, difference-in-differences, propensity scores, regression discontinuity, cluster randomization.

Causal effect of hybrid vs gasoline powertrain on fuel consumption (EPA data, 9 estimators + robustness suite)

TWFE plus Double Machine Learning on a synthetic 240k-row panel, with Optuna nuisance tuning under R-loss and DR-loss. Shows why the nonlinear basis must be built before entity demeaning.

Does swapping EconML's auto first stage for CatBoost improve CATE in Double Machine Learning? Same CausalForestDML second stage across 13 datasets, paired Wilcoxon. Auto wins PEHE on 7 of 13; tuned CatBoost wins on smooth DGPs.

Causal inference — ATE/ATT & propensity (IPW/AIPW) for DuckDB (Python)

Reproducible benchmark of uplift-modeling approaches (meta-learners, causal forests, DML, IV) on the Criteo Uplift dataset.