About

The thread through my work is failure: take a method people rely on, then work out exactly when and why it breaks. My TMLR paper does that for the R-learner in network causal inference, my SAM reproduction does it for a popular optimizer, and the bit-collapse project does it for post-training quantization.

I build things from scratch before trusting a library: an autodiff engine, SAM, a Hessian tool for loss landscapes. Right now I’m applying the same habit to privacy at WSAI, deriving the noise calibration for a new DP mechanism myself instead of treating it as a recipe.