automation
01Agentic pipelines and workflow orchestration that run unattended, including LLM-powered systems that do real work while you sleep. Human-in-the-loop design puts people exactly where judgment matters and nowhere else.
matrices systems that decide.
then adapt. then improve themselves.
matrices is the independent practice of sam cialek: automation, optimization, modeling, and applied AI. Fifteen years across quantitative finance, data science, and software engineering, pointed at one job: turning decisions you make by hand into systems that make them better than you do.
automationoptimizationmodelingaiautomationoptimizationmodelingaiautomationoptimizationmodelingai
01 / practice
what gets built here
Four disciplines, one point of view: a decision made more than once deserves a system, and a system deserves a model of its own uncertainty.
optimization
02Portfolio and resource-allocation optimization, plus decision analysis under uncertainty. Objectives are built on the full distribution of outcomes, not just the expected value, because the tails are where plans die.
modeling
03Bayesian inference in Stan, PyMC, and NumPyro. Causal DAGs and treatment-effect estimation, agent-based simulation with synthetic populations, and probabilistic forecasting that states its uncertainty out loud.
ai
04Applied LLM systems with real evaluation and scoring pipelines behind them, adaptive question engines, and machine learning that actually ships. If it cannot be measured, it does not get deployed.
02 / work
selected work
Three systems built end to end by the practice: a causal engine, an adaptive quiz, and prediction-market tooling.
serif
A consumer-health causal inference engine. It turns wearable and biomarker data into personalized, dose-aware recommendations through a causal DAG and Bayesian posteriors, so advice comes with a magnitude, not just a direction.
- 1,188 participants
- 3 cohorts
- causal DAG + Bayesian posteriors
prism
A Bayesian adaptive quiz that maps political disposition instead of sorting people into two bins. Questions adapt in real time to what the model still does not know, and converge in about 30 questions.
- 14 latent dimensions
- 130 archetypes
- converges in ~30 questions
markets
Tooling for prediction markets: Bayesian decomposition of contract prices into shared latent terms, so related contracts can be priced against each other instead of in isolation. Live dashboards and an active forecasting track record.
- shared latent-term decomposition
- live dashboards
- active forecasting track record
03 / credentials
the paper trail
B.A. Economics & Mathematics
NYU / 2011
MBA
Columbia Business School / 2019
M.S. Computer Science
Georgia Tech / 2024
Plus 15 years across quantitative finance, data science, and software engineering.
04 / contact
start with a conversation.
Thirty minutes, no deck required. Bring the decision you keep making by hand, and leave with a sense of what a system for it would look like.