A browser-native Bayesian inference engine and causal tree discovery tool. Making sophisticated statistics accessible without dumbing them down.

TypeScript WebAssembly Bayesian Causal Inference
Launch App → View on GitHub
⚡

Interactive demo loads here

Beta-Binomial calculator with real-time posterior updates

What Makes Tyche Different

Most A/B testing tools make you think backwards—they tell you the probability of seeing your data if there's no effect. Tyche flips this: it tells you the probability of success given your data. Pure Bayesian inference, no p-values, just honest probabilities about what you actually care about.

But that's just the start. Tyche is the only tool combining automatic heterogeneous treatment effect (HTE) discovery via causal trees with browser-native Bayesian inference. Find which customer segments actually benefit from your changes, not just average effects that hide the real story.

🎯 Conjugate Priors

Instant analytical solutions for conversion rates via Beta-Binomial updates. No MCMC needed for standard problems.

🌳 Causal Trees

Automatic discovery of heterogeneous treatment effects. Find which segments respond differently without fishing.

⚡ Variational Inference

Sophisticated VI engine for mixture models and complex posteriors. WebGPU acceleration coming soon.

🔒 Privacy First

Everything runs client-side. Your data never leaves your browser. No tracking, no uploads, no surveillance.

📱 Progressive Disclosure

Complexity revealed as needed. Start simple, dive deep when required. Mobile-first, touch-friendly.

🚀 Zero Dependencies

No servers to maintain, no APIs to pay for. Works offline, deploys instantly, scales infinitely.

Technical Architecture

Current Implementation

Active Experimental

Tyche uses a layered architecture optimizing for both performance and accessibility:

  • TypeScript Core: Type-safe statistical computing with compile-time correctness checks
  • Analytical Engine: Conjugate prior updates for Beta-Binomial, Normal-Normal, and Gamma-Poisson
  • Variational Bayes: Sophisticated VI for mixture model weights and complex posteriors
  • WebWorkers: Non-blocking computation keeping the UI responsive during inference
  • Progressive Enhancement: Works without JS for basic calculations, enhances with interactivity
// Example: Conjugate update for conversion rate
const prior = new BetaDistribution(1, 1); // Uniform prior
const posterior = prior.update({
  successes: 47,
  failures: 53
});
// Instant analytical posterior, no sampling needed
console.log(posterior.credibleInterval(0.95));

The Causal Tree Innovation

While everyone else does post-hoc segmentation (which leads to multiple testing problems), Tyche implements honest causal trees that automatically discover heterogeneous treatment effects while maintaining statistical validity.

This isn't just segmentation—it's principled discovery of where your treatment actually works differently, with proper uncertainty quantification at each leaf.

Why I'm Building This

Tyche is fundamentally a playground where I explore several intersecting interests:

  • Statistical API Design: What should modern statistical tools feel like? How do we make Bayesian inference as accessible as jQuery made DOM manipulation?
  • Browser-Native Computing: Pushing the boundaries of what's possible without servers. WASM, WebWorkers, and soon WebGPU for serious computation.
  • Accessible Mathematics: Can we make sophisticated causal inference approachable without dumbing it down? Progressive disclosure, not simplification.
  • Privacy-Preserving Analytics: Proving you don't need to upload data to analyze it. Your laptop is more powerful than you think.

Every design decision stems from a simple question: what if statistical tools were built for humans who make decisions, not statisticians who prove theorems?

Current Status

Beta Breaking Changes Expected

Tyche is under active development. The core Bayesian engine is stable and tested, causal trees are working but being refined, and the API is still evolving based on user feedback. If you're comfortable with beta software and interested in shaping the future of statistical tools, I'd love your input.

Try Tyche Now → Report Issues Send Feedback

Technical Documentation

For the technically curious, I'm documenting the journey: