I am a computer science PhD candidate at Stanford's Center for AI Safety. I study how language models can produce reasoning and memory that improve their own learning, and how reasoning systems organize knowledge. My work has moved from formal verification and interactive theorem proving toward AI alignment, reinforcement learning, and self-directed learning.

Website and publications

This page is a curated map. Repositories not listed here may be research notebooks, prototypes, old coursework, forks, or tools made for one particular machine.

• Markovian Transformers for Informative Language Modeling — Code, evaluations, and artifacts for our ICLR 2026 paper. The project studies reinforcement-learning objectives that reward chains of thought for being useful context for the model's own answer.

• Epistemic phase transitions in AI-era mathematics — A reproducible extension of my earlier work on proof-dependency networks to AI-assisted Lean corpora, including audited derived data and an interactive report.

• Weft — A tree-based interface for exploring many continuations from base language models, with token-level log-probability views and a second model that interprets what each continuation commits the text to.

• HessianAnalysis — An exploratory measurement framework for asking whether independently initialized language models converge toward similar gradient, function-update, subspace, and curvature geometry during training.

• ouroboros — A constrained self-modification experiment: an Emacs Lisp program can rewrite both its source and the prompt that determines how it will next be rewritten, while a small external kernel preserves the experimental record.

• UniversalBackrooms — A multi-model recreation of the Backrooms experiment, with configurable conversational templates and example transcripts. It is the most widely used repository on this account, but its model and API assumptions are now historical.

• Multifractal Visualizer — A GPU-accelerated OpenGL experiment for moving between fractal systems under deterministic and probabilistic update rules.

• Complete research artifact: MarkovianTraining accompanies the published paper, poster, slides, evaluation code, and experimental results.

• Active research: ept-ai-analysis is an audited, reproducible analysis with a live report; follow-up work on proof grain and full unfolding is still in progress.

• Active prototype: Weft has working GUI, terminal, and library interfaces and a passing local test suite.

• Research prototype: HessianAnalysis has functional training and plotting tools plus documented metrics, but not yet benchmark results, a paper, or a stable API.

• Early experiment: ouroboros has a documented kernel, published lineage, and explicit containment model; its scientific value is still exploratory.

• Historical but runnable: UniversalBackrooms is preserved as an experiment and collection of transcripts rather than advertised as a current multi-provider library.

• Visual prototype: multifractal-visualizer is a working research sketch with build documentation, not a supported cross-platform application.

• Exploring the Space of Short Programs — Patterns in very short programs built from the universal iota combinator.

• ManipulateProofTrees — Tools for extracting and analyzing the tree structure of Coq proof objects.

• What Is Music? — An older computational attempt to characterize harmony from first principles.

I use four status words deliberately: complete means a finished research artifact, active means I am still changing the work, prototype means the idea is clearer than the package, and historical means the repository is preserved but not maintained.

My publication rule is that a public repository must be safe to expose and possible for a stranger to interpret. A repository is featured only when it also has a useful entry point, an honest status, and some evidence that it works. Personal data, machine-specific state, and unreleased research stay private; superseded public work is archived rather than silently hidden.