Research
My research focuses on AI systems that explicitly represent causal structure and uncertainty. I am particularly interested in settings where predictive accuracy alone is insufficient because the real question is what will happen after an intervention or when the data-generating process changes.
Causal Discovery
Causal discovery aims to recover plausible causal structure from observational data. My work combines statistical discovery procedures with temporal structure, domain context, and language-model reasoning.
Agentic causal discovery
I investigate multi-agent workflows in which specialized agents propose, augment, and critique causal structure. Statistical conditional-independence evidence remains central, while contextual reasoning is used to help orient or recover relationships that are difficult to identify from finite observational data alone.
Key themes:
- expert-in-the-loop causal orientation
- small language models for contextual reasoning
- agent-based critique and edge augmentation
- causal story reasoning
- evaluation using structural Hamming distance and F1
Causal Forecasting
Forecasting models often perform well while leaving the data-generating mechanism implicit. My work studies whether learned causal structure can improve feature selection, robustness, interpretation, and intervention reasoning.
CaReBayes
Regime-Aware Causal Bayesian Forecasting for Non-Stationary Time Series
CaReBayes models non-stationary systems as a collection of regimes with different temporal causal structures. A causal graph is estimated for each regime, Bayesian structural models are fit over the discovered parent sets, and new observations are classified according to their predictive compatibility with each regime.
The broader idea is that distribution shift may sometimes reflect a change in mechanism, not merely a change in coefficients.
Causal feature selection for forecasting
Another line of work studies whether causal discovery can identify compact predictor sets that improve both forecasting performance and interpretability, particularly in infectious-disease forecasting.
Causal Simulation and Agents
I am increasingly interested in using causal models as the structural layer inside simulation systems.
Learning probabilistic agent behavior through causal discovery
Rather than manually specifying every behavioral rule in an agent-based model, this project asks whether observational data can be used to learn:
- which variables directly influence an agent’s decisions,
- probabilistic structural functions over those relationships, and
- uncertainty in both individual and emergent system behavior.
The resulting simulator could then be evaluated not only on whether it reproduces observed data, but on whether it predicts responses to interventions.
Amortized Bayesian policy simulation
A related direction uses amortized Bayesian inference to infer latent simulator parameters quickly across repeated policy scenarios. This could make uncertainty-aware simulation practical when traditional Bayesian calibration would be too expensive to repeat for every intervention.
Research philosophy
I am most interested in methods that satisfy three properties at once: they should be mechanistic enough to support intervention, probabilistic enough to represent uncertainty, and practical enough to evaluate on real systems.