Causal modeling, probabilistic machine learning, and the occasional bit of formal philosophy
I’m based in Cambridge, MA, and build causal Bayesian models with deep machine-learned components to explain why something happened. Lately that means treating a language model as a causal model and asking which heads in the network drive a behavior. It also means treating a fishery as a hybrid dynamical system and asking whether a given catch quota was necessary, sufficient, or both for keeping the fish population above a sustainable threshold. And it means modeling how a zoning change ripples through a city’s housing market. Transformer weights, fish populations, and city council votes don’t seem to have much in common, except that all three come from systems that are high-dimensional, dynamic, and only partly observed. I do this as a research scientist at the Basis Research Institute (New York/Cambridge, MA). I’m a core contributor to ChiRho, our open-source causal probabilistic programming language, and I’m responsible for its explainable reasoning module (with tutorials on actual causality, on categorical variables, and on continuous variables in an SIR model), and its spin-offs and applications.
I hold a PhD in Logic and Philosophy of Mathematics from the University of Calgary, and have held positions at Ghent University and the University of Gdańsk, with visiting fellowships at Northeastern University, the University of Edinburgh, Trinity College Dublin, and the University of Bristol.
Recent
- A Computationally Feasible Framework for Causal Probabilistic Explanation (arXiv preprint, 2026). Probabilistic Causal Impact recasts explaining why something happened as an estimation problem on a probabilistic causal model, generalizing actual causality and Pearl’s probability of causation so it scales past toy examples to a real deployed system trained on millions of datapoints.
- Whom to pity, whom to scold? (Information Processing & Management, 2026). A six-month field experiment on Reddit: 3,440 users who regularly attacked others received AI-assisted counter-speech, either normative or empathetic, and we measured what each kind of message did to their later behaviour.
- Causal Reasoning in Language Models, a special issue of Information Processing & Management that I guest-edited with M. Ptaszyński and R. Rzepka.
- Exploring the Maximally Sensitive Priors (Philosophy of Science, 2026, online first), with N. Lewandowski. When there’s essentially no evidence to inform a Bayesian prior, the Maximum Sensitivity method picks the prior that doesn’t depend on getting lucky about where the true parameter actually falls; we show that measuring inaccuracy by KL divergence collapses that recommendation to the older Maximum Entropy rule, and estimate how fast the recommended prior sharpens as sample size grows.
Currently working on
- LLM interpretability at scale. Combining probabilistic causal impact tooling with LLM architecture to scale up interpretability work, moving from one-off probing toward systematic, causal accounts of what drives model behavior.
- Epidemiological model evaluation. I’m working on automated reliability diagnostics for epidemic forecasting models, which would evaluate factors that most existing tools skip, such as whether inferred parameters are biologically realistic, whether a run of zero case reports reflects a reporting gap rather than an actual absence of cases, and whether a model’s derived variables hold up against independent data.
- Probability on Trial. A book under contract with Oxford University Press, co-authored with Marcello Di Bello, on using causal models and higher-order probability to evaluate forensic evidence.
- Common Task Framework for scientific ML. Collaborating with the Common Task Framework team to improve the pipeline and the metrics behind shared benchmarks for scientific machine-learning methods (forecasting, state reconstruction, generalization, and control on canonical dynamical systems). The team is launching a $50,000 competition on Kaggle, and I’m helping set up its technical side.
- PCI (Probabilistic Causal Impact). A causal attribution method, described in the paper and worked out in the accompanying computations, that decomposes explanations into separate necessity and sufficiency components instead of collapsing them. It bridges Halpern–Pearl actual causality with scalable attribution machinery (SHAP, LIME, gradient-based methods) by re-casting explanation as estimation on an expanded probabilistic causal model.
- Spatio-temporal causal modeling. Causal Bayesian models with deep machine-learned components for spatio-temporal valuation problems, dynamical models of policy interventions, and causal models of group foraging behavior in animals. The shared question is how to make causal explanation tractable when the system is high-dimensional, dynamical, and only partly observed.
Open source
- ChiRho ★ 277.
Basis’s open-source causal probabilistic programming language, built on Pyro/PyTorch.
I am a core contributor and lead its
Explainablemodule. - collab-creatures. Lead developer; the Bayesian and causal modeling toolkit behind our group-foraging research (Scientific Reports, 2024).
Most of my engineering work happens in private repositories. The chart below summarizes my GitHub activity by year: pull requests authored and merged (top), reviews and issues (bottom), across public and private repositories.
GitHub activity per year: pull requests authored and merged (top); reviews and issues (bottom). Public and private repositories combined; 2026 is year-to-date.