Demo

Try PokerAlpha

Start with the three demos below. They run on this site with nothing to install. The research explorers show precomputed engine output, and a few tools need the engine on your own machine.

Live in-browser demos
Research explorers

Look inside the trained strategy

Precomputed output of the real engine, stamped with the commit that produced it. These run on the public site.

Decision surface81 engine analyses over bet size and stack depth, showing where the recommendation flips.
Trust map169 preflop hands × 3 nodes with gate verdicts, visits and cross-seed disagreement.
Solver microscopeOne information set followed across training checkpoints and seeds.
Policy atlas6,000 stored policies projected to two dimensions.
Red teamAutomated search for fragile states, with an independent confirmation run.
Hand replayOne decision with the modelled opponent range, street by street.
Agent arenaSeat-swapped agent tournaments with confidence intervals.
Game labKuhn, Leduc and an auction duel on the same solver stack, with exact exploitability.
Local engine features

Tools that need the engine on your machine

Live analysis requires the PokerAlpha engine running locally (python -m poker_alpha.api), and it only accepts requests from a copy of this site served on your own machine. On the public site these pages show what is listed in the last column, not live analysis. How to run it locally.

toolwith the local engineon this site
AnalyzerRuns your own hand state through the full decision pipeline.bundled example spots
Strategy explorerBrowses the release strategy's preflop matrix node by node.bundled preflop nodes
Session reviewReviews an uploaded hand history after the session.a notice; uploads need the engine
TrainerAsks you to play a spot and grades you against the engine's mix.bundled demo spots
Screenshots

What the application looks like

Captures of the current pages and the desktop overlay, not mockups.

Watch a solver learn: After 10,000 CFR iterations in the browser. The curve is exact exploitability, recomputed as it trains.
Watch a solver learn. After 10,000 CFR iterations in the browser. The curve is exact exploitability, recomputed as it trains.
Arena: Kuhn poker against the CFR policy. Every hand is dealt and played in the browser.
Arena. Kuhn poker against the CFR policy. Every hand is dealt and played in the browser.
Phase II results: Abstraction refinement against capacity-matched controls, from exact river subgames.
Phase II results. Abstraction refinement against capacity-matched controls, from exact river subgames.
Engineering: How the C++ training backend went from 21.9 to 1,625 iterations per second.
Engineering. How the C++ training backend went from 21.9 to 1,625 iterations per second.
Desktop overlay: The Electron overlay built from source, showing demo data. Desktop builds are not released yet.
Desktop overlay. The Electron overlay built from source, showing demo data. Desktop builds are not released yet.

New here? The ten-step tour walks through the project in order.