Research

My research lives at the intersection of biomedical optics, inverse problems, and modern deep learning. The unifying goal is making functional neuroimaging usable at the neonatal bedside. I push aggressively across multiple threads in parallel — 13+ active research and tooling projects maintained simultaneously, with new commits almost every day.

Diffuse Optical Tomography for neonates

DOT reconstructs hemodynamic concentration changes (HbO and HbR) from diffuse near-infrared light measurements. The neonatal head is a special case: the skull is thin enough for light to penetrate cortical tissue, but the imaging geometry is small, motion is constant, and the inverse problem is severely ill-posed.

My thesis line of work — see piglet-realtime — develops a real-time multi-wavelength reconstruction pipeline validated on a piglet neonatal model. Key components include a Q-prior for ill-posed inversion and GCV-based automatic regularization parameter selection.

Differentiable photon transport

Classical Monte Carlo photon transport (MCX) treats the forward model as a black box. Differentiable Monte Carlo opens the door to end-to-end DOT optimization — backpropagating through the simulator to learn optode placement, source profiles, or learned regularizers. See diff-photon-transport.

Generative priors for inverse problems

Score-based and diffusion models have transformed image priors in MRI and CT reconstruction. I am adapting these approaches to DOT, where the inverse problem is more nonlinear and the data are far sparser. See diffusion-priors-dot and dot-test-time-compute.

Contrast-agnostic preclinical segmentation

Reliable DOT requires accurate anatomical priors. Preclinical piglet MRI comes in many contrasts; rather than retraining per-contrast, I use SynthSeg-style domain randomization and Segment Anything (SAM2) to build contrast-agnostic segmenters. See synpig (Dice 0.9622) and sam-synthseg-piglet.

Mechanistic interpretability of biosignal models

Tiny transformers trained on fNIRS and other biosignals are starting to outperform handcrafted features. What features do they learn? biosignal-interp trains sparse autoencoders on a small fNIRS transformer to ask whether the latent features map onto anything humans recognize.

Open research infrastructure

Most research time is spent on plumbing. iu-hpc-agent is a verified reference for AI coding agents on the Indiana University Big Red 200 and Quartz HPC clusters. research-os is an arXiv watcher with a DSPy classifier and Obsidian writer on cron. 3brown1blue is a Manim skill for generating 3Blue1Brown-style explainer videos from papers.