I’m a PhD candidate advised by Steve Zdancewic and Konrad Kording. I am generally interested in programming languages as tools for building and understanding neural networks.

I study how to inject algorithms into neural networks—to improve data and parameter efficiency, and to make them easier to interpret and control. Specifically, I explore:

⊢ Compilers that inject algorithms into neural networks
⊨ Proofs that these compilers are correct
✓ Applications in training, interpreting, and controlling neural networks

I also founded REPL, an NSF undergraduate research internship in programming language theory. It is designed to encourage historically marginalized groups to pursue PhDs.

Research

Preprints and Manuscripts

Compiling λ-calculus into LLMs— In preparation
Joey Velez-Ginorio, Nada Amin, Konrad Kording, Steve Zdancewic

Compiling to transformers — In submission
Joey Velez-Ginorio

Representative Publications

Talks

A past, present, and future for programming neurons · ‘26
Compiling to neurons · ‘26
Compiling to recurrent neurons · ‘26
Compiling to linear neurons · ‘26
When does X implement Y? · ‘24
Compiling to transformers · ‘24
When do neurons* represent True? · ‘23
Compilers as linking hypotheses · ‘22
A neural compiler· ‘21
Finding programs in the brain· ‘20
Compiler synthesis· ‘19
Interpreting actions by attributing compositional desires· ‘17