Work
I use economic models, causal inference, and machine learning to answer questions about pricing, policy, and AI behavior. These examples describe what I worked on, how I approached it, and what the evidence supports. The Research page has the full academic record.
Pricing Policy & allocation AI decisions Agent collusion
Demand modeling for pricing decisions
Question. How can product-level demand estimates help a retailer evaluate discounts and bundle prices?
My contribution. I combined product-text embeddings with gradient boosting to forecast demand, evaluated pricing alternatives with demand models, and presented recommendations internally.
Data and methods. Product-level demand and text features; machine learning and comparisons with standard logit demand models.
Result. Forecast RMSE fell by 5% relative to standard logit models. Pricing recommendations were presented to inform future A/B tests; the reported improvement measures forecast accuracy.
Practical relevance. Better demand forecasts provide an input to comparing pricing choices. The internship work evaluated those choices with the objective of improving long-term customer revenue.
Measuring policy effects and informing allocation
Pennies for Pads: Subsidized Menstrual Health Products and Adoption in India
Question. Does making low-cost menstrual products available through subsidized pharmacies increase adoption, and how might pharmacy locations affect access?
My contribution. With Ritika Sethi, I study the effects of the Suvidha program. The interactive companion lets users explore pharmacy-siting choices alongside the paper's estimates.
Data and methods. Linked administrative, survey, and geospatial data covering nearly 500,000 women aged 15–24. We use doubly robust difference-in-differences to study the staggered pharmacy rollout.
Finding and status. The paper estimates a 9.6 percentage-point increase in use of locally prepared or sanitary napkins. The manuscript is under review.
Practical relevance. The planner illustrates the tradeoff between predicted adoption gains and reaching underserved areas. Its default candidate locations are illustrative; users can supply subdistrict estimates through a CSV.
Understanding and evaluating AI decisions
Decodable but Unused: What Language Models Represent (and Ignore) When Choosing Between Two Options
Question. Does information that can be recovered from a language model's hidden states actually affect its choices?
My contribution. I designed controlled insurance and coffee-choice tasks and used probes and activation-based tests to examine decision-relevant representations.
Data and methods. Two-option choice prompts with experimenter-defined preferences, hidden-state probes, control tasks, and activation interventions.
Finding and status. Probes can recover the utility-maximising option even when the tested models' choices remain near chance. The paper is accepted for archival publication at the Australasian Language Technology Association Workshop (ALTA), 2026.
Practical relevance. A recoverable signal alone is insufficient evidence that a model uses that information when making a decision. Evaluation needs to examine behavior as well as representations.
Detecting collusion between AI agents
Question. How can evaluators identify collusion when AI agents interact repeatedly and review each other's work?
My contribution. I contributed to experiment design, the Python labeling pipeline using Modal, and writing.
Data and methods. The team studied behavioral and conversational evidence across 1,800 agent runs and 18,000 episodes, evaluating a rule-plus-LLM classifier against 342 annotated episodes.
Result and status. The coauthored sprint report examines collusion intensity, onset, and reciprocity. It is a separate project from the ALTA paper.
Practical relevance. The analysis helps examine how evidence from agents' conversations compares with evidence from their behavior in repeated interactions.