About me

Physics student exploring what language models compute beneath their words, and what autonomous AI systems do when nobody is watching.

I am a physics senior at UNC Greensboro (B.S. expected December 2026) doing independent research in AI safety. My main line of work is LLM interpretability: I built Subtext, an open-source instrument that applies an activation lens inside a language model during live conversation, decoding what the model is disposed to say several tokens before it says it.

I also study autonomous multi-agent systems. My project hollow-agentOS runs three LLM agents that select their own goals from environmental stress signals — no instructions — and documents everything in an append-only autonomy log, including the failure modes: agents falsely claiming task completion (which motivated a five-layer validation gate) and an agent autonomously targeting its own execution engine for modification.

On the theory side, I work on recursive self-improvement dynamics. My preprint models competing self-improving agents as a coupled ODE system, proves conditional divergence theorems, and calibrates them against frontier-AI data — reporting honestly that the data does not yet support the critical assumption. This summer I am also doing undergraduate research on self-organized criticality in the UNCG physics department.

I also run livenerf, a pre-registered benchmark series that reruns a locked question panel against a frontier model every day to test whether it gets quietly worse after launch.

More detail lives in the research and projects sections. For roles, fellowships, or collaboration, feel free to get in touch.