About me
Welcome. I am a PhD candidate in Economics at Harvard University, with a PhD minor in Data Science. I am fortunate to be advised by Prof. Michael Lingzhi Li and Prof. Jesse Shapiro. I do applied machine learning research at the intersection of AI, healthcare, and economics.
I received my BSc in Resource Economics from Renmin University of China. I also spent summer 2026 at Roblox as a Data Science Intern.
Research
My research sits at the intersection of machine learning, large language models, healthcare, and economics. I build AI tools for consequential decisions, and I use the tools of economics to measure what those decisions are worth.
- AI and healthcare. I combine machine learning and causal inference on large clinical datasets to measure the value of care delivery, such as how primary care affects patient outcomes after an ICU stay. I also work on data-driven methods for decision-making under uncertainty in healthcare operations.
- Multi-agent LLM systems. I am interested in how groups of language-model agents can work together effectively.
- Machine learning for social science. Earlier, I used vision transformers and multimodal contrastive learning to link noisy historical records at scale. I have also worked on empirical questions in economics, such as how crowdfunding backers respond to being pivotal and how outsourcing affects public auctions.
News
-
I will present The Value of Primary Care Beyond Prevention: Evidence from Intensive Care Unit Mortality at the INFORMS Annual Meeting 2026 in San Francisco, in the session Innovations in Healthcare: AI, Analytics, and Human Factors (Monday, November 2, 11:00 AM, Moscone South, Room 104). If you are attending, please feel free to reach out to grab a coffee and chat about research.
-
Joined Roblox as a Data Science Intern for the summer.
-
Quantifying Character Similarity with Vision Transformers appeared in the Proceedings of EMNLP 2023.
Papers
AI and Healthcare
-
Working Paper 2026INFORMS 2026
The Value of Primary Care Beyond Prevention: Evidence from Intensive Care Unit Mortality
TBD
Machine Learning for Social Science
-
EMNLP 2023
Quantifying Character Similarity with Vision Transformers
Record linkage is a bedrock of quantitative social science, as analyses often require linking data from multiple, noisy sources. Off-the-shelf string matching methods are widely used, as they are straightforward and cheap to implement and scale. Not all character substitutions are equally probable, and for some settings there are widely used handcrafted lists denoting which string substitutions are more likely, that improve the accuracy of string matching. However, such lists do not exist for many settings, skewing research with linked datasets towards a few high-resource contexts that are not representative of the diversity of human societies. This study develops an extensible way to measure character substitution costs for OCR'ed documents, by employing large-scale self-supervised training of vision transformers (ViT) with augmented digital fonts. For each language written with the CJK script, we contrastively learn a metric space where different augmentations of the same character are represented nearby. In this space, homoglyphic characters, those with similar appearance such as “O” and “0”, have similar vector representations. Using the cosine distance between characters' representations as the substitution cost in an edit distance matching algorithm significantly improves record linkage compared to other widely used string matching methods, as OCR errors tend to be homoglyphic in nature. Homoglyphs can plausibly capture character visual similarity across any script, including low-resource settings. We illustrate this by creating homoglyph sets for 3,000 year old ancient Chinese characters, which are highly pictorial. Fascinatingly, a ViT is able to capture relationships in how different abstract concepts were conceptualized by ancient societies, that have been noted in the archaeological literature.
-
arXiv 2023
Linking Representations with Multimodal Contrastive Learning
Many applications require linking individuals, firms, or locations across datasets. Most widely used methods, especially in social science, do not employ deep learning, with record linkage commonly approached using string matching techniques. Moreover, existing methods do not exploit the inherently multimodal nature of documents. In historical record linkage applications, documents are typically noisily transcribed by optical character recognition (OCR). Linkage with just OCR'ed texts may fail due to noise, whereas linkage with just image crops may also fail because vision models lack language understanding (e.g., of abbreviations or other different ways of writing firm names). To leverage multimodal learning, this study develops CLIPPINGS (Contrastively LInking Pooled Pre-trained Embeddings). CLIPPINGS aligns symmetric vision and language bi-encoders, through contrastive language-image pre-training on document images and their corresponding OCR'ed texts. It then contrastively learns a metric space where the pooled image-text embedding for a given instance is close to embeddings in the same class (e.g., the same firm or location) and distant from embeddings of a different class. Data are linked by treating linkage as a nearest-neighbor retrieval problem with the multimodal embeddings. CLIPPINGS outperforms widely used string matching methods by a wide margin in linking mid-20th century Japanese firms across financial documents.
Economics
-
Working Paper 2025
Does My Contribution Matter? Pivotality and Investment in Crowdfunding
We track the funding of over 7,000 Kickstarter campaigns, which succeed only if pledges reach a posted goal by a deadline. Two weeks before the deadline, projects just below their goal attract 1.2 percentage points more new backers per day than those just above it, in every crawl-batch sample and specification we examine; the one-week estimate, 1.4 points, depends on one crawl batch. The average pledge per backer does not move. A reward-seeking backer should prefer the just-funded project, so private returns cannot explain the drop. Neither can warm glow, the utility of giving itself. Standard altruism at the pivotal margin may: a backer who values success gains the pledge’s effect on success probability times that value. We call this pivotal altruism. Three and four weeks out the discontinuity is indistinguishable from zero. A calibration under a pre-registered protocol finds that its timing requires a response that does not scale with the pledges a campaign still needs: coarse beliefs that sort campaigns into at risk and safe reproduce it, rational Poisson beliefs cannot, and the model fails one of its three consistency checks. In that model, the goal maximizing expected funds collected is about 1.0 to 1.4 times what a campaign would raise without the pivotal push.
-
Working Paper 2025
Efficiency Improvement through Outsourcing and Incentivizing: Evidence from Judicial Auctions in China
This paper examines the efficiency of public service outsourcing in debt contract enforcement. Using the foreclosure auction outsourcing reform in China as a natural experiment, we find that outsourcing significantly increases the success rate of foreclosure auctions, driven by reductions in search frictions and transaction frictions. If all foreclosure auctions in China between 2016 and 2023 had been outsourced, it would have generated a total social welfare gain of 0.54 billion CNY and an additional 26.16 billion CNY in benefits for debtors and creditors.
Experience
San Mateo, CA. Worked on recommendations for new marketplace items.
Teaching
Teaching Fellow, Harvard University
- Econ 1126: Quantitative Methods in Economics, with Prof. Davide Viviano
- Econ 70: Personal Finance, with Prof. John Campbell