cd /news/artificial-intelligence/i-build-an-efficient-and-lean-poker-… · home topics artificial-intelligence article
[ARTICLE · art-139024] src=github.com ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

I build an efficient and lean Poker bot with AI

Developer aeneassoft released QuantPlay, a 6-max poker trainer whose Python engine (~1.5 MB) runs locally in the browser via WebAssembly and Pyodide, after a three-month attempt to rank in the top 5 of the GTO Wizard AI leaderboard fell short. The trainer offers six modes — GTO, Exploit, Arena, Punishment, Tournament and Match — against profiled bots, with bot decisions made locally and no LLM involved in play; measured in Node plus Pyodide, import takes 0.9 s, five full graded hands run in 0.32 s, and the slowest single request is 0.11 s (0.35 s in tournament mode). The heads-up bot Prince lost against opponents measured over 2,992 paired decks, and the UI is in German.

read6 min views1 publishedSep 24, 2026
I build an efficient and lean Poker bot with AI
Image: Michielbdejong (auto-discovered)

Play it now: quantplay.io — no account, no server, nothing to install. The whole engine (Python, ~1.5 MB) is downloaded once and runs on your machine via WebAssembly. Your hands never leave your browser.

QuantPlay grew out of a three-month attempt to build a heads-up bot that could rank in the top 5 of the GTO Wizard AI leaderboard. That goal was not reached (the honest balance sheet is !_PROJEKT_BILANZ_2026-09-10.md). What came out of it, and what this repository is about, is a 6-max trainer you cannot buy anywhere, a bot with every number measured and sourced, and a measurement discipline that refuted more ideas than it confirmed.

Six modes, all against a league of profiled bots (TAG, LAG, nit, station, maniac, whale …) whose play was tuned in paired self-play and gated against regressions:

Mode What it trains
GTO The bots play their baseline straight. Every decision you make is graded against the engine's own line within the hand, with a written explanation.
Exploit The bots learn you live and attack your leaks. You experience your own exploitability.
Arena Rotating, adaptive opponent types across all stack depths. A stress test for staying disciplined.
Punishment Five hunters, each built from a leak that was actually measured in human play (sheriff, iso-hammer, value press, trap nit, blind fighter).
Tournament A 60-player MTT (6 tables × 10) with rising blinds, antes, table balancing, a final table and a top-9 payout. ICM hints appear once the bubble factor bites. Exact Malmuth-Harville ICM, tested against independent enumeration.
Match Same opponents, no coaching, no distractions. Everything is recorded and graded for review afterwards.

Plus: pre-fold while others still act (the hand is played out in the background and chips move correctly), hand replay with per-decision grades, an opponent panel showing what the bots have learned about you, a session analysis after ~100 hands, and a 61-entry poker glossary whose formulas are the audited ones in knowledge_base/math/formulas.py. The UI is in German.

pip install -r requirements.txt
python -m pokerbot.web.six_server --open --trainer

Windows / PowerShell, Python 3.12, run from the repo root. The trainer is at http://127.0.0.1:8000/training. Bot decisions are local, instant and free; no LLM is involved in play.

web/ turns the same trainer into a static site. web/build.py zips the Python package plus the knowledge files it reads at runtime; web/src/worker.js boots Pyodide in a Web Worker, installs five pure-Python wheels and imports the trainer; web/src/bridge.js replaces fetch('/api/…') so the unchanged front-end talks to the worker instead of a server. The route dispatcher is pokerbot/web/browser_bridge.py, which calls the FastAPI endpoints directly (Pyodide has no threads for the ASGI threadpool). Tested under CPython in tests/test_browser_bridge.py. Measured in Node + Pyodide: import 0.9 s, five full hands with grading in 0.32 s, slowest single request 0.11 s (tournament mode 0.35 s).

python web/build.py                          # -> web/dist (deterministic, content-hashed)
python -m http.server 8765 --directory web/dist

The site is deployed from web/ to Vercel. Prince (the heads-up bot) does not take over heads-up pots in the trainer: measured in 2,992 paired decks it lost against the league core, so it is off by default and the browser build ships without its neural advisors.

A heuristic engine (preflop blueprint from CFR push/fold + solver-distilled tables, postflop equity / pot odds / MDF with solver-frequency advisors) wrapped in a chain of measured guards, a bounded exploit overlay, and optional real-time re-solving (TexasSolver) at river/turn nodes. Every number below has a source in docs/MESSKATALOG.md; the channel decides what a number means.

Measurement Result Channel / n Meaning
Heads-up vs GTO Wizard AI (the only true GTO anchor) −21.1 ± 9.4 bb/100 (v4, AIVAT) live API, n = 979 95 % band ≈ [−39.5, −2.7]. Leaderboard top: private bots at −3.1; best frontier LLM −9.2. The current champion (v5) was never anchored.
v5 vs its own base +30.6 bb/100 paired self-play mirror A non-regression bound, not strength. Individual guard gains (~+53) did not add up.
Tournament risk premium (proportional bubble factor) +10.0 ± 5.0 pp ROI paired SNG arena, n = 1,500 The full bubble factor was refuted (−8 pp); the proportional one validated.
vs PokerSnowie via the screen bridge +3.2 bb/100 [−37, +44] 3,114 clean hands of 3,651 Break-even; the automation errors, not the bot, cost the account.
Kaggle Game Arena heads-up (LLM field, 100 bb) −0.7 ± 2.5 bb/100 vs champion paired, 150 decks A cheap volume channel, not a GTO anchor.
Opponents who are not GTO +300 … +700 bb/100 local benchmark bots The exploit layer works against exploitable play.

What was learned the hard way is in the balance sheet: guards around a heuristic engine hit an asymptote (≈ −10 predicted, never reached); self-play gains do not transfer to a re-solver; the per-hand standard deviation is 294 bb, so a ±4 bb/100 answer costs ~5,400 hands; and a well-prompted frontier LLM plays heads-up better than this bot. Ideas that were refuted by measurement are listed explicitly in docs/MODULKATALOG.md — for anyone rebuilding, that is the most valuable part.

Path What
pokerbot/engine/ cards, evaluator (treys), Monte-Carlo equity, the N-player table (2–10 seats, side pots)
pokerbot/strategy/ preflop blueprint, range tracker, postflop math, advisors, exploit model, ICM, tournament doctrine, the guard chain ( auslese.py )
pokerbot/arena/ the opponent league ( sixmax.py ), MTT director, tournament arena
pokerbot/web/ the trainer server ( six_server.py ), the UI (static/training.html ), the browser bridge
pokerbot/coach/ decision capture, grading oracle, feedback templates, replay, opponent panel, glossary
pokerbot/autogym/ the self-improving loop: math oracle, paired gyms, gates, journal
pokerbot/benchmark/ GTO Wizard, Slumbot, Kaggle Game Arena harnesses, duplicate/paired evaluation
pokerbot/vision/ the PokerSnowie screen bridge (template matching, state gates, marathon guard)
knowledge_base/ extracted, structured knowledge: audited formulas, ranges, concepts, exploit playbook
docs/ STATE.md (live state),MODULKATALOG.md (272 modules),MESSKATALOG.md (every measurement), plans and consults
web/ the static browser build of the trainer (quantplay.io)
tests/ python -m tests.test_table ,test_bot ,test_icm ,test_tournament ,test_prefold ,test_browser_bridge

CLAUDE.md holds the working conventions for AI-assisted sessions, including the measurement doctrine (paired decks, A/A null test must be exactly 0, pre-registered expectations, three-run rule).

PolyForm Noncommercial 1.0.0. You may use, copy, modify and share everything here for noncommercial purposes: personal study, research, teaching, hobby projects, and use by noncommercial organizations. Any commercial use — selling, running as a paid service, using the bot, trainer, ranges, knowledge base or measurements inside a commercial product or to make money at the tables for a business — requires written permission. Ask via a GitHub issue or the contact on the organization page. Keep the NOTICE file (the Required Notice: line) with every copy.

Closed on 2026-09-10, reopened only to publish. Large artifacts (solver caches, trained nets, LLM checkpoints, hand histories) are not in the repository. The books the knowledge base was extracted from are not included either.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @quantplay 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/i-build-an-efficient…] indexed:0 read:6min 2026-09-24 ·