I work on evals and data for agentic coding at Google DeepMind.
Before that I was Employee #7 at Mechanize, where I worked on the most difficult long-horizon software engineering environments for frontier labs. I also led the development of GBA Eval, an OSS long-horizon coding benchmark — you can check it out.
I previously studied mathematics at Stanford and worked on empirical economics research in innovation and patents, primarily using language models as a measurement instrument. I also competed nationally in mathematics and policy debate competitions.
Mechanize· Senior Software Engineer
RL environments and evaluation infrastructure for coding agents.
Prior research at the intersection of machine learning and economics. Full abstracts on the papers page.
Yang, S. (2025). Understanding Innovation Quality and Success with Large-Language Models. Working paper. I use LLM embeddings of patent text to better understand innovative quality.
Cong, L. W., & Yang, S. (2025). Understanding Patenting Disparities via Causal Human+Machine Learning. Working paper. We develop a novel causal inference framework using ML and apply it to study USPTO decisions.
Yang, B., & Yang, S. (2025). Endogenous Fracturing under Partisan Voting. Working paper. A model of legislative coalitions; applied to the McCarthy speakership.
Yang, S. (2024). New Criteria for Triangle Similarity. Mathematics Magazine. Analytical proof of two new triangle similarity criteria involving angle ratios.
© 2026 Stephen Qingyuan Yang