Building AI systems that learn and reason about causation
I develop causal AI systems that learn structure, model how systems change, and reason about the consequences of interventions under uncertainty. My work brings together causal discovery, forecasting, Bayesian inference, simulation, and agent-based modelling, with a growing focus on causal world models and physical AI.
Research areas
Causal Discovery
Learning causal structure from observational data, statistical tests, and contextual knowledge.
Causal Forecasting
Using causal structure to improve forecasting robustness, interpretability, and intervention reasoning.
Simulation & Agents
Combining causal models, Bayesian inference, and agent-based simulation for policy and complex-system analysis.
Selected work
Regime-aware causal Bayesian forecasting
A forecasting framework for non-stationary time series that learns regime-specific temporal causal structure and combines it with Bayesian prediction, uncertainty quantification, and intervention reasoning.
Causal feature selection for infectious-disease forecasting
A causal feature-selection approach for identifying compact, interpretable predictor sets that can improve machine-learning forecasting while reducing reliance on purely associational features.
Agentic causal discovery
My recent work explores causal discovery systems in which statistical evidence is complemented by small language models and specialized agents for orientation, augmentation, critique, and causal-story reasoning.
View research → · Read the papers: Agent-guided causal discovery, FLLM 2025 → · Causal story-guided discovery, FLLM 2026 →
Current research direction
Much of my current work focuses on causal AI, particularly the intersection of causal discovery, forecasting, simulation, and decision-making.
I am especially interested in how these ideas can contribute to world models and physical AI. Many AI systems can be viewed as learning a policy
\[\begin{equation} \pi_\theta(a_t \mid s_t), \end{equation}\]
which determines what action an agent should take in its current state. World models go a step further by learning how the environment evolves under those actions,
\[\begin{equation} T_\phi(s_{t+1} \mid s_t, a_t). \end{equation}\]
I am particularly interested in giving these learned world models an explicitly causal representation,
\[\begin{equation} T_\phi(s_{t+1} \mid s_t, \operatorname{do}(a_t)), \end{equation}\]
where actions are treated as interventions on the environment. The broader goal is to develop AI systems that can reason about possible consequences before acting, while remaining robust, interpretable, and uncertainty-aware.
To see more details (especially if the notation is unfamiliar) learn more about my current research interests.