Developed methods to interpret all neurons in Llama3-8B, explain model failures like "9.11 > 9.8", and steer models away from reasoning failures. Also developed Docent, a tool for analyzing and understanding agent traces, and Predictive Concept Decoders, a new end-to-end interpretability architecture
AI Researcher at Imbue
2023-2024
Scaled up LLM pretraining and infrastructure. Trained reward models to improve code reasoning.
Software Engineer Intern and Social Media Manager at Windsurf
Summer 2022
Worked on improving the efficiency of ML inference workloads. Interfaced with C++
bindings for popular ML frameworks, implemented compiler optimizations and language features,
and developed infrastructure for model quantization. Also managed the company Twitter account
Zero Knowledge Researcher at 0xPARC
Spring 2022
Implemented and optimized the first ever implementation of the optimal Ate pairing
within a zk-SNARK, and then wrote servers to verify proofs of BLS signature validity
MIT Web.Lab Instructor
Winter 2021, Winter 2022
Taught a web development class to ~300 MIT students. Prepared lectures, advised
projects, debugged code. MongoDB + Express + React + NodeJS Stack.
Researcher at PRIMES-USA
Spring 2018 - Summer 2019
Predicted the movements of refugees across a region during
refugee crises. Devised new models that work as accurately as existing
models and more efficiently. Recognized as a Regeneron STS Finalist, 9th
Overall. Published in the IMA Journal of Applied Mathematics, Volume 85, Issue 6, December 2020
Selected Awards
International Math Olympiad (2017-2019): Represented the USA, individually ranked 103rd (Silver), 10th (Gold), 11th (Gold)