AI safety research, infrastructure, and BioML. Currently a research intern at EleutherAI (SOAR), working on cheap evaluation methods for inoculation prompting. In terms of engineering I build data pipelines, backend services, and ML systems for AI safety and BioML research.

San Diego, CA · LinkedIn · Email

Inoculation prompting and off-target effects — Developing an open, low-cost method to detect when inoculation prompts also suppress unintended traits or change model personas. Supported by grantmaking.ai and a Thinking Machines Lab Tinker Research Grant.

UmamiBench — Building a benchmark for scientific judgment in frontier LLMs, including task design, scoring rubrics, and the model evaluation pipeline.

Positional encodings in vision transformers — Investigating the mechanisms and effects of absolute and rotary positional encodings through mechanistic interpretability.

RecA — Developing ML methods to prioritize RecA inhibitors as potential antibiotic adjuvants, in collaboration with researchers in the University of Cambridge Department of Genetics. Built the computational stack: compound curation, potency classifiers, and a docking pipeline for millions of compounds, using 40,000+ CPU-hours on GCP.

Some current safety work isn't public yet.

Selected merged pull requests from @OpenCodingSociety:

Mechanistic interpretability and evaluation · Empirical Alignment and Pretraining · ML for biology and scientific imaging · Data pipelines and backend infrastructure