Projects

Modular pipeline for eviction risk modeling with strict temporal separation.

Budgeting Discretion

Under Review Ongoing

This research formalizes street-level discretion as a finite-budget resource allocation problem, proving that optimal decision-makers follow a dynamic threshold rule determined by the tail-heaviness of potential welfare gains. By analyzing years of homelessness service data, the study demonstrates that human caseworkers strategically ration their discretionary authority in response to real-world capacity openings and operational timing. The findings highlight that discretion is a managed resource, offering a unit-free framework for comparing decision-making personalities across different institutional settings.

Fixed Points and Stochastic Meritocracies

Upcoming FAccT '26

Meritocratic selection into scarce opportunities (e.g., elite schools, jobs, or housing) is often treated as a best-case mechanism for fairness: choose the strongest candidates today, and do so symmetrically across groups. This project shows why that intuition can fail over time. We study an inter-generational selection process in which admission increases future “merit,” so that today’s selection changes tomorrow’s applicant pool. Even when two groups start identical and the per-period rule is deterministic and merit-based, stochastic shocks can create substantial transient gaps. In an Equal-Advantage benchmark, these gaps eventually wash out and the system converges back to parity, with the speed of convergence depending on program capacity and efficacy. But when we introduce a minimal feedback loop—an Affinity-Advantage in which the currently leading group’s non-admitted members receive even a tiny boost—random early leads become self-reinforcing, producing persistent long-run separation and, beyond a threshold, extreme dominance. Simulations in a richer continuous-ability model mirror the theory: small feedback advantages can entrench inequality despite individually fair decision rules. The results emphasize that static fairness at a single decision point is insufficient in dynamic settings, and that policy and algorithm design must explicitly account for long-run feedback, scarcity, and the stability properties of the induced process.

Street-Level AI

published AIES '25

This project investigates the reliability of Large Language Models (LLMs) in "street-level" bureaucratic roles, specifically within the high-stakes domain of homelessness resource allocation. By comparing LLM-generated prioritizations against real-world data from St. Louis, the research demonstrates that current off-the-shelf models exhibit significant internal inconsistency and fail to align with established, socially determined vulnerability scoring systems. The study reveals that while LLMs can mimic lay human judgments in simple pairwise tests, they struggle to replicate the nuanced, context-sensitive discretion of professional caseworkers, suggesting that "vibe prioritization" is not yet ready for unmediated integration into critical social safety net decisions.

Discretionary Trees

Published AAAI '24

Street-level bureaucrats (e.g., homelessness caseworkers) must implement policy while retaining discretion to make exceptions in individual cases—an essential feature that can improve outcomes but also raise concerns about bias and procedural justice. Using administrative records from the St. Louis HMIS (2007–2014) during a period when intervention assignment was less formulaic, this project uses machine learning to make caseworker decision-making legible: (i) how much of it can be captured by simple, human-usable rules, (ii) whether decisions remain consistent beyond those simple rules, and (iii) how discretion is applied and with what apparent consequences. Short, interpretable decision trees capture a substantial portion of assignments, suggesting plausible “default” heuristics, while higher-capacity models show that caseworkers are highly consistent overall. We operationalize discretion as systematic deviations from the short-tree baseline and show that these deviations are not random: discretionary decisions disproportionately target lower-vulnerability households, and discretionary upgrades to more intensive interventions align with higher expected marginal benefit—evidence consistent with strategic, welfare-improving discretion rather than arbitrary overrides.

Two-Sided Bandits in Matching Markets

Preprint

We study repeated two-sided matching when *both* sides must learn preferences over time (a “two-sided bandits” setting) and agents cannot explicitly communicate. Prior approaches typically assume that the receiving side (“arms”) has fixed, known preferences (often even common knowledge), which makes it possible for proposers to avoid conflicts and converge to stable outcomes. This project develops decentralized algorithms that provably converge to stable matchings in strictly harder regimes: (i) when arms’ preferences are known to arms but *not* to players (private), and (ii) when arms themselves are also uncertain about their preferences and must learn them from interaction. Our key idea is to combine optimistic beliefs about (a) reward/value estimates and (b) the probability a proposal will be accepted, so players learn while matching rather than in a separate “explore then commit” phase. The resulting methods (OCA-UCB for private arm preferences and PCA-SCA for fully unknown preferences) converge to stability and drive regret down in simulation, with Thompson-style belief tracking often converging faster and more smoothly than UCB-style tracking in the fully-unknown setting.

Increasing the Value of Information in Online POMDP Planning

Published Undergrad Thesis

Online POMDP planners like POMCP can perform near-optimally on many problems, but they systematically undervalue information-gathering actions when there is a long delay between collecting information and using it. This thesis identifies that failure mode and introduces a lightweight modification to online planning that better reflects the value of information in such settings: add an entropy-based term to the UCB1 action-selection heuristic in POMCP. The resulting algorithm (POMCPe) biases search toward trajectories that meaningfully reduce belief uncertainty, without sacrificing the anytime nature of Monte Carlo planning. Using a new benchmark variant of the hallway problem designed to amplify long-horizon information delays, the thesis shows that entropy-augmented planning substantially improves performance and more reliably chooses the early information-gathering detour that disambiguates the environment before committing to high-stakes decisions.