To inhabit the world of creations and actions inasmuch as I do the world of words...
Nicholas Thompson Birthday Run
2/2026I ran a guerilla marketing campaign to get writer and editor Nicholas Thompson to come to my birthday run.
4/2024For my undergrad thesis I studied algorithmic fairness, a field with roots in theories of justice and motivated by real-world case studies. I explored a framework for "differential expressiveness", a method for evaluating bias based on the idea of the same feature meaning different things for different groups.
A Hierarchical Model for Degree-Heterogenous Random Graphs
12/2023A teammate and I formulated a new model for generating random graphs whose nodes have varying degree by considering a setup where nodes have different "popularities" drawn from some latent distribution. We outlined methodologies for performing inference and simulating graph generation using our model.
Analyzing CoT Prompting via Gradient-Based Feature Attributions
7/2023Before reasoning models existed, my team inspected the then-nascent technique of chain-of-thought prompting using gradient-based techniques (calculating "saliency scores" based on the gradient of a language model's prediction). Workshopped at ICML 2023.
Weighted Consecutive-Loss Rules for Justified Representation
1/2023Given an electorate that votes repeatedly, how can we define fairness in terms of who's elected, and how should votes be weighted in order to satisfy those criteria? A collaborator and I investigated several voting "rules", proved theoretical properties about them, and simulated how they work.
Cluster-Robust Standard Errors
12/2022My team studied cluster-robust regression, a flavor of statistical inference that assumes data can fall into clumps within which the standard errors from fitting a linear model may not be uncorrelated. We summarized estimation techniques for more accurate error modeling in these cases, replete with R demo code.
5/2022With the Data to Actionable Knowledge lab at Harvard's SEAS, I explored the behavior of toy neural networks in continual learning on binary classification tasks, and tested the effectiveness of the Bayes Factor in non-nested model selection, where it's hard to compare the relative complexity of two models.
Surveying Differential Privacy in Regression Methods
4/2022Differential privacy is a branch of computer science that analyzes sensitive data in a way that preserves anonymity. The challenge is doing this in a high-fidelity way, so that we get statistical results that aren't too noisy. I implemented some techniques for performing linear regression under this framework.