Hi!
I'm Noah! I recently graduated from Brown University where I studied computers and the many things you can do with them.
I'm spending a year as a Sutter Hill Ventures Codepoint Fellow; currently working on formal verification for hardware at SiFive.
I'm interested in
- Compilers, Program Analysis, Solvers, Theorem Provers. I think compilers and solvers are some of the most powerful/flexible tools for all sorts of problems.
- AI and Mechanistic Interpretability. I think it's kind of bonkers that we have powerful models performing sophisticated algorithms yet we only understand a small portion of them. I'm interested in the line of research that aims to characterize model computation and reverse engineer it.
- Cognitive Science. Understanding how humans think and make decisions the way they do has been a long-time interest of mine.
I started a blog recently to think better thoughts and develop stronger opinions.
Previously, I...
Recent Projects
See resume for others
TLang: a low-level representation for transformer circuits
Inspired by previous work that represents and compiles simple programs to transformer weights, I prototyped a low-level representation for transformer circuits. It differentiates itself by expressing arbitrary transformer weights, representing the continuous nature of transformer computation more faithfully, and supporting RoPE.
Dimspector: PyTorch Static Shape Inference
A devtool that statically infers the shape of PyTorch tensors to provide inlay hints and catch shape mismatches immediately. Still in development.
ATA
ATA is a student-facing chatbot that acts as a tutor I built with a few friends for use in Brown courses. Professors drop in their course materials and get course analytics; students can then chat with ATA about lectures and assignments. ATA is university-sanctioned and in use by hundreds of students across multiple courses at Brown. We are lucky to be supported by funding from B* and API credits from OpenAI. ATA was also featured in the 2025 edition of Brown CS's magazine, Conduit.