independent practice · automation · optimization · modeling · ai
the glyph is an integral. every unknown gets a distribution, and the distributions get integrated out.
systems that decide, adapt, and improve themselves.
matrices works where four disciplines overlap: systems that hold their beliefs as probability distributions, update them against evidence, and get better with every pass.
Agentic pipelines and workflow orchestration that run unattended and know when to stop. Every automated decision carries its own uncertainty, so handing off to a human is a designed event, not a failure mode. People sit exactly where judgment matters, and nowhere else.
Allocation, scheduling, and decision analysis under uncertainty. Objectives are scored against the full posterior distribution of outcomes rather than a point estimate, because the tails are where plans break. Marginalizing over what cannot be known is not a compromise; it is the method.
Bayesian inference is the house discipline. The model is written down in Stan before the first data point is trusted, and the assumed causal structure is drawn as a DAG so every arrow is on the page and open to argument. When a problem wants to live in Python end to end, the same model moves to PyMC or NumPyro without changing what it believes. Nothing ships until posterior predictive checks show the fitted model can generate data that looks like the world it claims to describe.
model {
effect ~ normal(0, 0.5); // prior: skepticism, quantified
sigma ~ exponential(1);
y ~ normal(x * effect, sigma); // likelihood: the data gets its say
}Applied LLM systems with evaluation built in from the first commit: scoring pipelines, adaptive question engines, and machine learning that ships. The interesting failures get measured, not retold as anecdotes.
consumer health
A consumer-health causal inference engine. Wearable and biomarker data flow through a causal DAG, and Bayesian posteriors turn them into personalized, dose-aware recommendations. Built across 3 cohorts.
political disposition
A Bayesian adaptive quiz over 14 latent dimensions. Each answer updates the posterior, the next question is chosen for what it would reveal, and the engine converges on one of 130 archetypes in about 30 questions.
prediction markets
Prediction-market tooling that decomposes contract prices, Bayesianly, into the latent terms they share, surfaced on live dashboards.
Fifteen years across quantitative finance, data science, and software engineering. The constant throughout: decisions under uncertainty, treated carefully enough to build systems around them.
A short call is the fastest way to find out whether a problem is a fit for the practice.