I'm Matthew Reda. I build Bayesian measurement systems for marketing and finance decisions. These are the fifteen repositories I worked on between May and November 2026.
2,287 commits across 15 repositories
Marketing mix modeling that commits to its assumptions before it sees the fit.
Read the documentation API reference
The likelihood, priors and identification strategy are fixed before the fit, which removes the room to shop for a result.
Hierarchical effects share strength across geographies and products without flattening the differences between them.
Mediated pathways such as media to awareness to sales, joint KPIs with correlated errors, and halo and cannibalization between products.
A standalone Bayesian engine with a FastAPI service and a React studio. Versions up to 1.5.0 remain on PyPI under Apache-2.0; later versions are not publicly distributed.
Fit, score and share labelled multi-dimensional models, safely.
Open the docs Source on GitHub
A model is a declarative YAML spec or a small Python plugin. Data is validated against an xarray schema, and the fitted preprocessing is replayed exactly at score time.
Fits and scores run in isolation. Whole analyses run as versioned workflows, from unit-root tests to rolling-origin backtests against naive forecasts.
A governance policy requires a model to beat itself fitted on shuffled data and to give no credit to fake noise drivers. A second person signs off.
Declare what you want to know. Find out whether the data can tell you.
Bayesian causal decision science in one vocabulary: identification, experimental design, calibration against randomized results, response surfaces and meta-analysis. Checked against 21 published numbers from four real datasets, with twelve narrated examples from agronomy to product analytics.
Pass a model between tools without losing what the numbers mean.
A draft standard for exchanging models, test results, budget scenarios and decisions. Every exchanged number carries its estimand, gates degrade before they block, and records land in an append-only log.
Illustrative record
Real data, a real decision, and a first model that breaks.
Fifteen challenges and more than a hundred worked examples on PyMC 6, ArviZ 1 and PyTensor 3. Every challenge ends in a decision someone has to make, and the obvious first model fails in a way you have to diagnose. Hints are tiered, opt-in and tracked.
Every chart computes in your browser, from data with a known truth.
A scroll-driven explainer of causal identification, double machine learning and doubly robust ad-lift estimation. It is one self-contained page, and the Monte Carlo simulations run in a Web Worker.
Why a model that predicts well can still mislead you about what to do.
An interactive guide to causal identification: a nine-stage practitioner workflow, a design router, a geo-test sizing calculator and a field guide to twelve graph types. A companion page answers 51 practitioner questions with sources.
The reading behind the work, kept in the open.
A research vault of more than 200 notes on Bayesian statistics, econometrics, causal inference and discovery, agent-based modeling and market response models. The explainers on this page cite it note by note.
Upload a survey. See how exposed respondents compare with comparable controls.
Fits a propensity-score model, adjusts the control group by matching or weighting, and reports lift for each brand KPI with bootstrapped uncertainty. It also shows where the two groups overlap and where they do not, feature by feature.
A weekly early read on whether media is still working.
A leading indicator built from a marketing mix model's prediction errors, with a tracker app for planners, data scientists and the analysts who build it. A KPI-only mode fits its own model when there is no MMM. Tested with a simulation study and a robustness battery.
What is one more brand study worth?
A value-of-information simulation for brand measurement. It simulates worlds with long-memory awareness, geo surveys, a national tracker, brand-lift reads, platform experiments and a TV geo holdout, fits a joint model at each rung of evidence, and scores the result by budget-reallocation regret.
Survey noise, treated as survey noise.
State-space models for brand-tracker proportions with sampling error, linking media spend to latent brand to sales. The app runs the naive analyses on the same simulated data and shows what they get wrong.
Operator tools for a social media app.
Client work for Epoch Lag, where I build the ML pipelines, the admin dashboard and the infrastructure underneath them.
Parametric parts from a versioned component library.
build123d models that share a library of patterns, mechanisms and fasteners. One command builds a part, checks printability, exports STL, STEP and 3MF, renders it and compares its metrics with a golden record.
This site.
An Astro blog with 29 essays on measurement, from adstock identification to sequential stopping, and the page you are reading.
Every repository with a commit between 11 May and 11 November 2026, ordered by activity.
| Repository | What it is | Built with | Commits | Last commit | Visibility |
|---|---|---|---|---|---|
| modeling-api | Platform for labelled multi-dimensional models | Python, xarray, PyMC, TypeScript | 1,215 | 19 Oct | Public |
| mmm-framework | Bayesian marketing mix modeling toolkit | PyMC 6, FastAPI, React | 423 | 11 Nov | Private |
| axiom | Bayesian causal decision science library | Python, NumPyro, JAX | 172 | 28 Aug | Public |
| ump | Draft standard for exchanging models and decisions | JSON Schema, Pydantic, TypeScript | 167 | 1 Nov | Private |
| pymc-challenges | Real-data Bayesian modeling challenges | PyMC 6, ArviZ 1, Jupyter | 79 | 6 Nov | Public |
| redam94.github.io | Blog and portfolio | Astro, TypeScript | 66 | 10 Nov | Public |
| printing | Parametric 3D-printable parts | Python, build123d | 56 | 7 Oct | Public |
| second-brain | Public research vault | Markdown, Python | 41 | 9 Nov | Public |
| admin-dashboard | Admin dashboard and ML pipelines for a client app | Python, TypeScript, Terraform | 36 | 21 Oct | Private |
| double-ml-explainer | Interactive explainer of double machine learning | HTML, JavaScript | 15 | 10 Nov | Public |
| causal-vs-predictive-modeling | Interactive guide to causal identification | HTML, JavaScript | 7 | 10 Nov | Public |
| propensity-modeling | Propensity scoring and brand lift analyzer | Python, FastAPI, React | 4 | 5 Nov | Public |
| media-effectiveness-index | Weekly leading indicator of media effectiveness | Python, JavaScript | 4 | 26 Oct | Private |
| brand_voi_sim | Value-of-information simulation for brand studies | Python, NumPyro | 1 | 31 Oct | Private |
| latent_var | State-space models for brand-tracker surveys | PyMC, NumPyro, Streamlit | 1 | 16 Oct | Public |
Projects from before May 2026.
The model exists to support a decision. If there's no decision, there's no model, only homework.
The work doesn't start with a model. It starts with a half-hour conversation about what the business is actually trying to decide. Everything after that follows the same six steps.
Before any modeling, get clear on what's actually being decided: the budget that'll move, the threshold someone has to cross, the trade-off on the table. The model exists to support that.
Pick a likelihood, priors and identification strategy before you see the fit. Specification shopping is the fastest way to get an answer that flatters the brief and survives nothing.
Schema, coverage, zero-spend stretches, collinearity. Most modeling problems are upstream data problems. Most overconfident answers are too.
R̂ below 1.01, ESS above 400, no divergent transitions. If diagnostics fail, the summary statistic is fiction, and a fiction shipped to a CFO is worse than no answer.
Intervals, not point estimates. Sign uncertainty visible. "We don't know yet" is a complete sentence. Stakeholders adjust faster to honest ranges than to numbers that quietly move every quarter.
Posterior draws turn into expected ROI, the regret of cutting a channel, the probability a budget shift beats the status quo. If a chart doesn't change anyone's behavior, it's decoration.
MIT physics, then half a decade in measurement science. The throughline: take uncertainty seriously, write the model down first, ship code.
Built MMM and incrementality systems for portfolios in the high-eight-figure media-spend range. Owned methodology and fitting infrastructure.
Hierarchical regression for brand health. Started the migration from frequentist regressions to fully Bayesian posteriors.
Undergraduate research on Stochastic Optical Reconstruction Microscopy: single-molecule localization, point-spread-function fitting and Bayesian deconvolution.
I take a small number of engagements at a time. Each one starts with the same question: what is the business trying to decide, and what would honest measurement of it look like? The conversation is free.
Two weeks. I sit with your marketing, finance and analytics leaders, map the decisions on the table, and write the measurement plan that would credibly support each one. Often the most valuable output is the model we agree not to build.
Six to twelve weeks: schema, ingest, Bayesian model, decision layer. You get the repository, the service and an analyst trained to run it. Every output ties back to a decision the business made up front.
A quarterly retainer alongside your in-house analytics team: methodology calls, code review, and the difficult conversations with stakeholders about what the posterior says and what it doesn't.
Send a short note about the decision on the table and what currently supports it. I'll tell you whether honest measurement can move it, or whether you should save the budget and buy your team lunch instead.