I am on the 2026–27 academic job market.
About
I am a Postdoctoral Research Associate at Stanford University, in the Center for Sustainable Development and Global Competitiveness, working with Prof. Jie Wang and Prof. Michael Lepech. I was previously an Associate Research Scientist at Columbia University, and I received my Ph.D. in Computer Science from the University of Virginia in 2024, advised by Prof. Bradford Campbell.
I work on building reliable AI systems. My research follows three directions:
Trustworthy LLM Generation and Agents
NeurIPS ×4 · ICLR · ICML ×2 · AAAI ×2 · WWW · NeurIPS D&B · EMNLP Industry Best Paper Nom., etc.
More detailsHide detailsMultimodal and Multi-Domain Representation Learning
NeurIPS · ICML · KDD ×2 · ACM MM ×3 Oral (top 4%) · SIGIR · CVPR, etc.
More detailsHide detailsFederated and Efficient Learning
NPJ AI · ICCPS · IPDPS · IEEE TSC, etc.
More detailsHide detailsTrustworthy LLM Generation and Agents
Most AI systems that generate things are graded by a similarity score: close enough to a reference counts as correct. That is not the same as checking it. My approach is to make the output checkable: first recover the parts of the target that are fixed (the layout of a web page, the structure of a program, the anatomy in an X-ray), have the model generate within them, then verify the result against that structure and repair what fails. Where nothing can be run to check an answer, such as a medical question, I build the evaluator and study where it breaks. The line runs from front-end code, which has to compile and render, to radiology reports that must point at the right anatomy, to judges for text that cannot be executed, and to shared infrastructure for evaluating agents at scale.
Selected papers; the full list is on the Publications page.
Multimodal and Multi-Domain Representation Learning
Alignment here is not about preferences; it is about keeping different kinds of data in step inside one model. A recommendation model, for example, learns from user histories and from video content at the same time, serves a homepage and a live stream at once, and predicts quantities such as watch time that are not a single number. Standard training quietly assumes that all of these behave alike, and at scale they do not. My methods measure the gaps and correct for them during training. Large-scale recommendation has been the testing ground: the methods were tested in live A/B experiments and one is fully deployed.
Selected papers; the full list is on the Publications page.
Federated and Efficient Learning
Privacy and computing budgets are usually treated as constraints applied after a model is designed, which is why privacy protection often costs so much accuracy that it gets switched off. I decide at design time which parts of a model stay on the device, which parts are shared, and what sharing costs: a rule for where privacy noise is worth paying, a way to combine updates from devices that differ from one another, and a cost metric that says when an AI deployment has paid for itself. The line started in datacenters and now runs on phones and sensors, in buildings and in hospitals.
Selected papers; the full list is on the Publications page.
My long-term goal is Translational AI Systems: deploying these methods in healthcare, buildings, finance, software and recommendation, etc.
Translational AI Systems
Healthcare · Buildings, Energy & Infrastructure · Finance & Open Source · Software Engineering · Platforms & Media
Best Paper Nom. EMNLP Ind.'25 · Oral (13.3%) CHIL'23 · Best Paper Cand. BuildSys'24 · Oral (top 4%) ACM MM'25
More detailsHide detailsTranslational AI Systems
The three lines are put to the test with the people who run the systems: clinicians, building engineers, developers, platform teams, and the open-source finance community. Each paper is tagged with the line it advances.
Selected papers; the full list is on the Publications page. Trustworthy LLM GenerationMultimodal Representation LearningFederated and Efficient LearningSystems & open source
I build and test this work with clinicians, building engineers, developers and platform teams. My work appears at top-tier AI venues including ICML, NeurIPS, ICLR, AAAI, KDD, WWW, SIGIR, ACM MM and CVPR, in the ACL family, in the Nature Portfolio journal NPJ Artificial Intelligence, and at systems venues including ICCPS, IPDPS, SenSys and IPSN, and it has been recognized by the World's Top 2% Scientists list, a Google HE Faculty AI Fellowship, a Best Paper Nomination at the EMNLP Industry Track, orals at ACM MM and CHIL, and a Best Paper Candidate at ACM BuildSys. I also contribute to the FinRL open-source ecosystem and the Harbor agent-evaluation framework, and organize the FinRL Contests.