VLM interpretability
Diagnosing and mechanistically characterizing spatial reasoning failures in vision-language models: linear probes, activation patching, and readout-layer analysis across LLaVA, Qwen-VL, InternVL and friends.
About
I'm an astronomer turned machine-learning researcher based in Haiku, Maui. I did a Ph.D. at the University of Hawaiʻi (measuring the chemistry of distant galaxies from the light of red supergiant stars), then spent time in data science, venture capital, and a startup before landing back at telescopes, this time pointing them with neural networks.
Day to day I lead a machine-intelligence research lab at U.S. Space Systems Command, building autonomous sensing systems for space domain awareness. On my own time I'm working on interpretability of vision-language models: why they fail at spatial reasoning, and what's going on in their internals when they do.
Right now
Autonomous sky sensing
Spectroscopic, neuromorphic and multispectral ML systems, from raw photons to deployed telescope networks. Star streaks, satellites, and all-sky calibration without human help.
Selected papers
- Accepted · NeurIPS 2026 · Datasets & Benchmarks Sophon: A Procedural Diagnostic for Spatial Reasoning in Vision-Language Models
- Journal of the Astronautical Sciences 2026 Rapid Deployment, Calibration, and Training of Optical Observatories for Space Domain Awareness
- PASP 2026 SENPAI: Sidereal Enriched Rate-Track Astrometry in Deep Imagery of Solar System Bodies
- WACV 2022 SpectraNet: Learned Recognition of Artificial Satellites from High Contrast Spectroscopic Imagery
- ApJ 2015 Red Supergiants as Cosmic Abundance Probes: The Sculptor Galaxy NGC 300
- Advances in Astronomy 2012 Transit Analysis Package: An IDL GUI for Exoplanet Transit Photometry
2,200+ citations across 55+ publications. Full list on Google Scholar →
The lab
Open-source things I've built.
AllClear
Blind all-sky astrometric solver and directional sky-transparency mapper for autonomous observatories.
SENPAI
Zero-calibration, sensor-agnostic astrometry for detecting stars and satellites. pip install astro-senpai
SDASim
Speed-optimized, differentiable satellite scene simulator for synthetic training data. PyTorch.
Burr
Autonomous telescope calibration and observation scheduling for rapid-deploy observatories.
astroeasy
Containerized astrometry.net plate solving with a friendly Python API.
TAP
Transit Analysis Package: Bayesian MCMC exoplanet light-curve fitting, widely used in the exoplanet community.
Sophon
Procedural spatial-reasoning benchmark for VLMs. pip install sophon-bench
About the name
Ballistic Burrito: burritos when you want, where you are. Stuck in traffic? Drop your window. Our patented launch process perfectly melts your cheese in flight. 78.6% effective delivery rate, and climbing. Interested investors, please inquire.