PokerLab AI

For publishers and partners

A model of how a real poker pool plays, and the best response to it

What exists today, how it is checked, and what I will open to a partner who wants to look under the hood.

What the product does

PokerLab AI builds a statistical model of a specific player pool from hand histories and answers, for any spot and any hand, how that pool plays: which actions, how often, with which sizes. On top of the model it computes the best response against that pool, hand by hand, and shows where the value of the exploit comes from, street by street to showdown. It is not a solver and not a language model. A solver answers "how do I avoid losing to a perfect opponent"; this answers "how do these people actually play, and what is the most profitable reply".

What is built and working

Range ExplorerRanges of the player to act in any spot: preflop and postflop, with the frequency of every action per hand, sizing distributions and hand-strength mix behind each size.
Best responseExploit of the pool at flop, turn and river, with the value of the spot "as the field" against "with the exploit", and an EV report per hand that walks the tree to showdown.
Decision treeThe range split by the hand properties the pool really responds to on a given board, so "why does this hand fold while that one calls" has a visible answer.
Several gamesThree demonstration models at the moment: 6-max cash at two limits and a zoom 8-max game with a straddle, each trained with the same recipe on mined hand histories of its pool.
Expected error on screenEvery spot carries an estimate of how far the model is expected to be from the field there, measured on held-out hands.
Repeatable pipelineOne training recipe, applied to each game without manual tuning; every model package carries a manifest of what it was built from.
Fast analytical databaseA column-oriented (OLAP) store for mass data analysis of hand histories: it feeds the models and the hand reviews today, and is on its way to becoming a standalone analytics product.

How it is validated

Each model is checked on hands it never saw: a frozen held-out set, plus a second held-out set reserved for small differences. Two kinds of measurement: how well the model predicts the pool's action frequencies in every spot (visible for every decision), and how well it predicts actions given the actual hand (visible at showdown). A new version replaces the old one only when the primary metric improves and nothing else gets worse.

What I will open to a partner. The site is kept deliberately simple: the metrics and the inner workings are left off its pages so that a player sees the tool, not the laboratory behind it. To a publisher or partner evaluating the product I will provide, under a simple non-disclosure agreement, whatever is needed to form an opinion:

  • the details of the architecture and answers to any question about how it works;
  • all accuracy metrics per game and per street, with baselines beside them;
  • data volumes and the share of showdown hands per game, the held-out protocol and the version history of the model;
  • the cost of a model: compute hours and wall time from raw hand histories to a deployed package, measured, not estimated;
  • known weak spots: rare lines, deep stacks, multiway pots, the approximations used for the flop exploit and their measured cost;
  • live access to the explorer and, if useful, a model trained on a pool of the partner's choice from the partner's own hand histories.

Nothing in the list above is a promise of future work. It all exists and is written down; I am offering to show it.

Exploit: from one spot to the whole game

Today the best response is computed for the spot in front of you, with the value of the spot and a report of where that value comes from. Next comes the exploit map of a whole game: the best response over the entire model, ranked by how much each weakness of the pool is worth. Then the map goes into a trainer: play hands against the model of the pool, with the exploit as the opponent or as the coach beside you.

Stages

Status: built in progress next planned

  1. 1 · builtModel of the poolHow a real pool plays, in any spot.
  2. 2 · builtExplorer and best responseRanges, sizes, the exploit of a spot.
  3. 3 · in progressLessonsThe model's answers as rules a player can read.
  4. 4 · nextExploit map of the gameWhere the pool leaves the most money.
  5. 5 · plannedTrainerPlay against the model of the pool.
  6. 6 · plannedYour own poolModels from your hand histories.
  7. 7 · plannedAnalytics productThe database behind it all, on its own.

Where this is going

The models on the site are illustrations. The product is the pipeline behind them: a model trained on the hand histories a player or a team provides for their own game, for the whole pool or for a type of player within it, with player types found automatically when needed, and a library of already modelled games as the starting point. The next stages are a job queue and user accounts, billing for computation, and upload of a user's own hand histories. Stages, not dates; I will share the criteria for each.

What I bring and what I am open to

I bring a working method that no solver offers: a model of a real pool, a best response to it, and a pipeline that turns a stack of hand histories into such a model for any game. I am open to a publisher or partner with an audience of players or coaches who would put it to use: as a product under their brand, as a tool inside a training programme, or as a service for a pool of their choice. The form of the partnership is a conversation, not a condition.

Questions and a walkthrough on request: support@pokerlab.ai. See the tool itself: Range Explorer.