Research Scientist at Granica, working with Andrea Montanari on approximate leave-one-out methods for non-convex learners and on Large Tabular Models.
I’m a computational statistician with a background in causal inference and machine learning. Before joining Granica, I built generative-AI systems at Google and lectured graduate scientific-computing courses at Stanford ICME. My work sits at the boundary between rigorous statistical methodology and the engineering of large-scale models.
I’m currently focused on two questions: how to predict out-of-sample error in non-convex settings without retraining, and how to build foundation models for tabular data that respect column semantics.
Currently
Estimating LOO Hessians via low-rank updates of the full-sample Hessian, masking eigen-directions associated with negative curvature. Prediction error for single-index DGPs with mismatched Teacher–Student stays within single-digit percent.
Incorporating column semantics into foundation models for tabular data, and measuring dataset generalization via per-example permutation of columns.
Trajectory
Working directly with Chief of Science Andrea Montanari (Stanford EE / Statistics) on approximate LOO for non-convex learners and on Large Tabular Models.
Founding engineer on YouTube Shorts: built the reward model and probabilistic shelf-triggering policy that replaced the original deterministic placement, implemented in production C++.
Core contributor to the Gemini 1.5 Technical Report: designed a self-critique LLM evaluation framework with embedded rubrics, and proposed model-assisted estimation to combine human-rater labels with LLM scores into an unbiased, lower-variance estimator.
Most recently: implemented a Diffusion Transformer for video super-resolution, following CDDT, written in JAX as sole author of the prototype.
Taught graduate Python (CME 211) and C++ (CME 212) to ~200 students per year alongside my industry role. Managed staffs of 4+ TAs per quarter. Awarded Best Lecturer in three separate years by students and faculty.
Thesis with Guido Imbens (Nobel 2021): a quasi-experimental causal estimate of post-nightlife performance in NBA and MLB, validated against bookmaker spreads (arXiv). Distributed min-cut work with Reza Zadeh (draft); summer at Lawrence Livermore on ML-based sepsis prognosis with Kaiser Research (abstract).
Primary contributor to the R codebase supporting Daniel McFadden’s expert testimony on causal damages from the 2010 Gulf Oil Spill. Discrete-choice model over a 26,000-household / 40,000-destination travel survey; fiscal impacts mapped across all 3,000+ U.S. counties.
Honors thesis with Gregory Duncan: a quasi-experimental estimate of the causal effect of NCAA D1 athletic participation on GPA (PDF). Found a +0.8 GPA effect for football players with low entering SAT scores; the opposite sign (~−0.5) for women’s crew. Walked on to Cal Swimming & Diving.
Selected Work
Self-contained PyTorch codebase for pretraining transformer LLMs from scratch: RoPE/NoPE, MoE (shared and routed), KV caching, LoRA, mixed-precision and gradient accumulation. Config-driven scaling from small to billion-parameter models.
Core contributor on LLM-as-judge evaluation. Designed a self-critique framework with embedded rubrics achieving inter-rater reliability comparable to humans; proposed model-assisted estimation to combine human and model scores.
Quasi-experiment using bookmaker spreads as a conditioner. Teams visiting cities with higher nightlife indices consistently underperform expectations; effect replicates across NBA and MLB and dissipates with rest >24h.
Summer research at Lawrence Livermore National Laboratory with Kaiser Research: machine-learned and Bayesian models for sepsis trajectory and prognosis from EHR signals.
Honors thesis at U.C. Berkeley using walk-on entry and mid-career injury as within-subject identification. Heterogeneous treatment effects by sport and entering SAT.
Draft from a summer of work with Reza Zadeh at Stanford on a distributed algorithm for graph min-cut.