what i'm working on

I'm currently working broadly on LLM agent research:

With Professor Chao Zhang, I work on RL training for enhancing LLM agent reasoning and tool use. I also build the infrastructure behind this work, including sandbox management systems and asynchronous training frameworks.

With Professor Qirun Zhang, I work on using LLM agents to accelerate LLVM compiler bug fixing.

With Professor Celine Lin in the EIC Lab, I work on using LLM agents for hardware design.

select projects

• DuckTrack — a tool for recording human computer use to build training data for multimodal agents. github / blog post
• emergent misalignment — replication and investigation of emergent misalignment from benign fine-tuning. github / LessWrong
• Multiturn RL SWE — multiturn reinforcement learning for software engineering agents. github
• LLM from Scratch — a small LLM running inside MIT's Scratch, the block-based kids programming language. github
• PRM annotation platform — a platform for annotating process reward model training data. github

fun things about me

I've been hosting Minecraft servers for 13 years, which is how I originally got into Linux, system administration, and computers in general. These days I run infra for the GT Drehmal community, an RPG adventure Minecraft server that I founded, and help host events for the Cross Collegiate Minecraft League (2CML).

I'm a member of the AI Safety Initiative (AISI) club at GT as well as an organizer at DuckAI, an open source research community at GT.

I played piano for 10 years, and I particularly enjoy classical or new age music. I also played baritone horn.

I'm a three-stripe white belt in BJJ, beginner climber, and avid hiker.

I enjoy cooking (LLM-assisted) and origami!

interests