three systems, one thread.

Wearable and biomarker data in, dose-aware recommendations out, and a causal model in between that knows what it does not know.

problem

Wearables and blood panels produce a stream of numbers and almost no guidance about what to change. Generic advice ignores the only thing that matters: how your outcomes respond to your behaviors, at your doses.

approach

serif begins with a causal DAG: sleep and training load act on biomarkers, biomarkers act on outcomes, and every arrow is a commitment that can be checked, not a decoration. The model was written down in Stan before the first data point was trusted, and Bayesian posteriors are fit over the edges of the graph, so each recommendation carries a dose and an uncertainty, never just a direction.

theta[c]  ~ normal(mu_pop, tau);      // cohorts share strength
effect[n] ~ normal(dose_curve(theta[cohort[n]], dose[n]), sigma);

the update

When a new biomarker draw arrives, serif does not start over. The fresh draw enters as the likelihood, everything believed the day before stands as the prior, and their product, renormalized, is the posterior the next recommendation is built on.

result

The engine has run for 1,188 participants across 3 cohorts, turning raw wearable and biomarker streams into recommendations that are personal, dose-aware, and honest about what the data can and cannot support.

Political disposition as 14 latent dimensions and 130 archetypes, measured by a quiz that rewrites itself as you answer.

problem

Political labels compress a person onto one axis, and fixed-form quizzes spend most of their questions confirming what the first few already established. The interesting structure, how someone reasons and not just where they land, never gets measured.

approach

prism keeps a running posterior over 14 latent dimensions and matches it against 130 archetypes. Each next question is chosen for its expected information: the engine marginalizes over the answers you have not given yet and asks the one whose answer would move the posterior most. Nothing about the order is scripted; the quiz you take is assembled by your own answers.

result

The estimate converges in about 30 questions: enough to place a person among 130 archetypes without asking a hundred things their earlier answers already implied.

A prediction-market price is one number wearing many beliefs. The tooling here takes the number apart.

problem

Related contracts move together because they share underlying drivers, and the price alone will not tell you which driver moved. Trading or forecasting on raw prices means reasoning about a sum without seeing its parts.

approach

A Bayesian factor model, prototyped in PyMC while the structure was still in question and rewritten in NumPyro once speed mattered, decomposes contract prices into a few shared latent terms plus a residual each contract keeps to itself. Live dashboards hold the decomposition current as prices move, so a move in the headline number can be read as a move in a particular belief.

result

Live dashboards over the decomposition, and an active forecasting track record built on the same machinery: forecasts placed against the latent terms rather than against the headline number.