what i’m up to
I’m a fourth-year PhD candidate at MIT. I build large-scale orchestrated AI and optimization workflows for complex tasks.
things i’ve done
MIT (2023–2027) — PhD candidate with Prof. Patrick Jaillet. Working on optimizing large-scale agentic systems, with applications in solving open math problems and finance.
Stanford (2019–2023) — Honors degrees in Mathematics (B.S.) with Prof. Jonathan Luk and Philosophy (B.A.) with Prof. Michael Bratman. Undergraduate Honors thesis with Prof. Yinyu Ye. Journalist at Stanford Daily.
Meta (2025–now) — Research collaboration.
Amazon (2025, 2026) — PhD intern.
things i’ve built
ScaleMath A control and optimization system to scale up mathematical discovery. It has solved 200+ named open conjectures, primarily through counterexample construction, after 20k+ agent hours.
QuantamentalAI An agentic valuation system specializing in the pharmaceutical industry that forecasts FDA decisions, with a mechanism to control for look-ahead bias. [Dashboard]
places i’ve spoken
Shanghai Jiao Tong University (2025) | Berkeley Simons Institute (2023) | INFORMS Annual Meeting (2024, 2025, 2026) | INFORMS Optimization Society (2026) | INFORMS Applied Probability Society (2025) | ICML (2026)
what i think about
When One Request Becomes Many: Work-Stability Audits for Agentic LLM Workflows · coming soon
Which Version Did the Agent Use? Controlling Versioned LLM-Agent Workflows with Consumed Lineage · coming soon
Uncertainty Control for LLM-Parameterized Decision Pipelines · coming soon
Agentic Commerce: Robust Policy Audits under LLM Decision-Maker Shift · coming soon
Design of Large-Scale Agentic Workflow for Fundamental Analysis · coming soon With MIT LFE
Parameter-free BlockMDP · coming soon With Meta
When Price Bands Stabilize Black-Box Pricing Algorithms · coming soon
End-to-End Learning of Correlated Operating Reserve Requirements in Security-Constrained Economic Dispatch
A Two-Layer Framework for Joint Online Configuration Selection and Admission Control · ICML 2026
Choice-Model-Assisted Q-learning for Delayed-Feedback Revenue Management
A Single-Sample Polylogarithmic Regret Bound for Nonstationary Online Linear Programming