Motif of the Mind

Work

What I'm working on

Less a portfolio than a map of the threads I'm pulling on right now, and the questions underneath them. Most of the work happens in the open. If a thread interests you, come build.

rekursiv.ai

Autonomous science discovery with AI

Founding Scientist / Chief Epistemologist

We're building a cockpit for autonomous research: teams of AI Scientists that form hypotheses, design evals, run experiments, and review each other's work. Every claim is traced back to evidence.

Thinking with AI agents: knowledge as process

Can LLMs generate real knowledge? Yes. But explaining something to an agent, then correcting its answer, also surfaces your own tacit knowledge. Working with an agent is a form of teaching, and teaching is how you find out what you actually know. I'm building and writing in public to test that.

  • Creator

    tuningfork

    A sampler benchmark built around the garden of forking paths. It records the branches that fail, because the failure path is knowledge too.

    • Bayesian
    • benchmark
    • open-source

  • Creator

    agent-team

    A disciplined multi-agent workflow for developing the BlackJAX ecosystem: explicit roles, plus a worklog that keeps the threads, decisions, and lessons agents would otherwise lose.

    • agents
    • open-source

  • Creator

    memoires

    An expert-curated catalog of Bayesian modeling knowledge, distilled from years of community forums. Evidence is graded and contradictions stay visible, so an agent can consult the hard-won answers instead of deriving them again.

    • Bayesian
    • agents
    • open-source

  • Open-source contributor

    sagent

    Multi-provider agent CLI and Python library.

    • agents
    • open-source

Essays on this, as they land: see Writing.

Probabilistic programming & Bayesian computation

I work on making rigorous Bayesian inference composable, fast, and genuinely usable. The tools are built in the open, with a community I care about.

  • Sole developer & curator

    BlackJAX

    Fast, composable Bayesian inference in JAX: samplers as building blocks, plus a companion sampling-book of tutorials and worked recipes.

    • Bayesian
    • JAX
    • open-source

  • Core developer

    PyMC

    A leading probabilistic programming library in Python for Bayesian modeling and inference.

    • Bayesian
    • PPL
    • open-source

  • Contributor

    TensorFlow Probability

    Probabilistic reasoning and statistical analysis: contributions to tfp.mcmc.

    • Bayesian
    • MCMC
    • open-source

  • Co-author

    Bayesian Modeling and Computation in Python

    A hands-on book on Bayesian modeling and computation (Martin, Kumar, Lao; CRC Press, 2021).

    • book
    • Bayesian

Roots: cognitive science

Before Bayesian computation, I trained as a cognitive scientist. My PhD and postdoc focused on how culture shapes visual perception. Underneath the experiments were larger, less testable questions I never stopped turning over: how a mind models other minds, and how cognition and consciousness might arise from computation. They were closer to philosophy than experiment then, and still are. Working with LLMs and agents has brought me back to those questions, so I'm slowly writing them down (clearly marked as speculation).

From that era: iMap4, a toolbox for statistical fixation mapping of eye-movement data.

Speculative · not peer-reviewed · revisiting, not predicting.