independent practice · automation · optimization · modeling · ai

matrices

systems that decide, adapt, and improve themselves.

the practice of sam cialek book a consultation →

01the practice four disciplines, one loop

matrices works where automation, optimization, modeling, and ai overlap: systems that make decisions, measure their own performance, and get better with every pass.

Every engagement opens the same way: with the question written before the query. What decision will this system make, and what would it cost to be wrong? The causal structure is drawn as a DAG and argued over until it survives contact with the people who know the domain. Identification is argued before estimation, always; if an effect cannot be reached from the data by any path, no amount of fitting will conjure it.

Automation comes after the inference earns it. Pipelines run unattended because the model can tell when it is out of its depth, and stops; human judgment is placed exactly where it matters, and nowhere else.

Optimization closes the loop. Objectives are scored against the full posterior rather than a point estimate, because the tails are where plans break; the same posterior tells an adaptive system which question, dose, or trade is worth the next unit of attention.

02method one identity, applied carefully

One identity carries the whole practice.

p(θ | y)posterior ∝ p(y | θ)likelihood · p(θ)prior

(1)

normalized by the evidence, p(y | θ) p(θ) dθ : every value of θ that could have produced the data, integrated out. the tall glyph on this page is that integral.

The model is written down in Stan before the first data point is trusted: priors stated out loud, a likelihood that respects how the data came to exist, every assumption in one file where it can be attacked.

model {
  theta ~ normal(0, 1);      // the prior, stated out loud
  y ~ normal(theta, sigma);  // the data gets its say
}

listing 1 · the shape of every engagement, in a language that forces assumptions into the open

Fitting is deliberately boring, and that is the point: warmup, chains, diagnostics, nothing believed on arrival. Afterward the model is asked to imitate the data it was fit on, and the imitation is compared with what actually happened; the gaps are read like a diff, because that is what they are.

03selected work 3 projects

serif

1,188 participants

consumer health

A consumer-health causal inference engine. Wearable and biomarker data flow into a causal DAG, and what comes back is personal: dose-aware recommendations with Bayesian posteriors attached, built across 3 cohorts.

prism

130 archetypes

political disposition

A Bayesian adaptive quiz that locates political disposition on 14 latent dimensions. Every answer updates a posterior over the archetype space and decides which question is worth asking next; the engine converges in about 30 questions.

markets

live dashboards

prediction markets

Prediction-market tooling. Bayesian decomposition takes contract prices apart into the latent terms they share, so related markets can be read as one model; an active forecasting track record keeps the method honest.

04about est. nyc

sam cialek

Fifteen years across quantitative finance, data science, and software engineering. The setting kept changing; the job never did: decisions under uncertainty, treated carefully enough to build systems around them.

M.S. Computer Science georgia tech, 2024

MBA columbia business school, 2019

B.A. Economics & Mathematics nyu, 2011

05contact sam@matrices.earth

start a conversation.

A short call is the fastest way to find out whether a problem is a fit for the practice.