Stephen Mell

stephenmell@cmu.edu | Twitter | GitHub | Google Scholar | DLBP | OrcID

Postdoc in PL/ML/systems at CMU advised by Zhihao Jia. PhD from UPenn, advised by Osbert Bastani and Steve Zdancewic.

I am interested in the intersection of programming languages, artificial intelligence, and computer systems, with two main research directions:

Spec-level programming. As coding agents take over much of software development, verifying and maintaining the code they produce becomes a central challenge. My work raises the level of abstraction at which programs are written, disentangling high-level behavior from low-level performance optimizations so that concise but precise specifications, rather than sprawling codebases, become the object that humans and agents reason about. Correct-by-construction optimizations then recover performance without sacrificing clarity. For example, I am building a system that automatically async-ifies and streams simple sequential Python, yielding large speedups without the complexity of hand-written concurrent code.

Languages and frameworks for agentic systems. Existing frameworks for agentic workflows and harnesses are limited in expressiveness, and most optimizations target the underlying models in isolation, discarding valuable information from the surrounding program. I am developing a general-purpose, Pythonic framework for agentic systems, analogous to what PyTorch is for neural networks, that treats the whole system as a program to be optimized for accuracy, cost, and latency, while supporting debugging, observability, and deployment.

Research

Preprints

Conference Publications

Other Publications

† Ordered alphabetically

Invited Talks

Funding

Teaching

Service

Education

Emplyment