Projects

An end-to-end analysis on the public Dunnhumby retail dataset — NMF + K-Means segmentation feeding meta-learner / Causal-Forest HTE and an OPE-validated targeting policy.

A simulation study unifying Incremental·Shapley channel credit and a path-level decomposition on a single Inhomogeneous Poisson Process — answering channel budget, journey design, and population causal effect under one efficiency identity (18-method benchmark, ground-truth MAE 0.016,

The grounded RAG assistant on this site — a safe LLM on a static Cloudflare edge that answers only from the published notes. The demo is the button at the bottom-right of this page.

An OPE deployment gate that rules "trust / distrust / send it to A/B" from recommendation logs alone, under the real-world condition that nobody knows the true value $V(\pi_e)$ — plus a benchmark that scores those signals' forecasting power and in-principle blind spots on a truth-holding backstage.