I Built an AI Poker Dealer to Practice Push/Fold — Meet Maya

Home game players never get enough push/fold reps. Real money games happen once a week if you're lucky; short-stack tournament spots happen maybe twice a session. That's something like 50 meaningful push-or-fold decisions per year — not nearly enough to develop muscle memory. So I built one of our AI dealers, Maya, into a heads-up sit-n-go you can play in your browser. She uses Nash equilibrium ranges plus a little randomness so she's not a robot, and you get unlimited push/fold practice without losing real money to learn what a Nash chart already could've told you.

The problem with reading a push/fold chart

Open a Nash push/fold chart and you'll see a clean 13×13 grid: green cells for "push these hands," red for "fold these," organized by stack depth in big blinds. Three minutes of study and you can identify the right action in any pre-flop spot.

And then you sit down at your home game, the blinds are 100/200, you've got 1,800 chips, and you look at your hand: nine-eight suited. The chart's clear answer disappears under three layers of "wait, is 9 BB the same as 10 BB on the chart?" and "the chart said BB call but I'm in the small blind facing nothing, that's different, right?" and "I think 98s was green at this depth but I'm not sure." You take too long. You either push because you said you would, or fold because you panic.

Reading a chart and using a chart are two different skills. Reading is comprehension; using is recognition. You build recognition by doing the thing under noise, hundreds of times, until it becomes automatic. That's what Maya is for.

What the game actually is

A heads-up sit-n-go. Both players start at 1,500 chips, blinds 50/100 — 15 big blinds each, the canonical "now we're in push/fold territory" depth. Blinds don't escalate; the format keeps you in the Nash zone the whole game. One player busts, game ends, replay.

The only legal actions are push (all-in) and fold. No raises, no flat-calls, no bluffs with min-raise sizings. This is intentional. The point isn't to simulate every poker spot; it's to drill the one decision that matters most for tournament survival, until that decision is automatic. A real tournament at 15 BB has roughly two legal options anyway: shove or fold. The game just makes that explicit.

How Maya actually decides

For each hand, the game knows: Maya's hole cards, the effective stack depth in big blinds, and whether she's shoving (SB-first-to-act) or calling (responding to your shove). From those three things, she does a lookup against the published Nash equilibrium chart for heads-up play.

The chart works in two layers:

1. Hand strength ranking. Every starting hand has a Sklansky-Chubukov value — basically, the maximum stack at which open-shoving that hand is profitable against perfect calling. Pocket aces are nearly infinite; seven-deuce off is around 0.7. Maya knows the SC value for all 169 hand classes.
2. Stack-depth cutoff. At each stack depth, Nash specifies what percentage of hands should be in the shoving range (or the calling range, which is tighter). At 10 BB, the SB shove range is 37% of hands. At 15 BB, 22%. At 20 BB, 14%. Maya computes the cutoff and checks if her hand makes it.

If she stopped there she'd be a robot — perfect Nash for every spot, every time. So the last step: a small random deviation, plus or minus 15% on the threshold each hand. Sometimes she shoves slightly tighter than Nash, sometimes slightly looser. This makes her feel like a real opponent who's trying to play Nash but isn't perfect — which is exactly what a strong human opponent looks like. Crucially, the deviation is small enough that her decisions are nearly always correct, just not predictable.

What makes Maya a useful practice opponent isn't that she's unbeatable — it's that she's realistic. She makes Nash-correct decisions with a small amount of noise on the margins, the same way actual tournament players do.

The 7-card hand evaluator under the hood

Pre-flop is fun, but the showdown is where the math gets serious. When both players go all-in, the game deals the rest of the board (flop, turn, river), then evaluates each player's best 5-card hand from their 7 available cards (2 hole + 5 community).

There are 21 possible 5-card subsets of 7 cards; trying all of them and picking the best would work but be slow. Instead, the evaluator uses bitmask tricks: rank counts in an array, suit counts in another, a single bitmap of which ranks are present at all. From those three pieces of information it can determine every hand category — straight flush down through high card — in linear time without enumerating subsets. Each hand becomes a 24-bit integer score; higher score wins. Same engine powers the equity calculator at /equity.

The result: on every showdown, the game can deal a board, evaluate two 7-card hands, declare a winner, and animate the cards in under a millisecond. The bottleneck is the animation, not the math.

What you'll actually learn from playing it

About three things, in this order:

1. Pre-flop card recognition

You'll stop seeing "ace-jack" and start seeing "AJ offsuit at 11 BB in the BB facing a shove — that's a clear call." You'll stop seeing "low pair" and start seeing "fives at 8 BB, that's a shove in the SB but a fold facing villain's shove." The labels merge with the actions until they're a single piece of recognition instead of two separate steps.

2. The shape of the Nash range

After 100 hands you'll have internalized that Nash ranges are weirder than they look on paper. Suited cards punch above their rank. Pocket pairs hold value much later than you'd expect. King-rag offsuit is a fold longer than ace-rag offsuit because the ace blocks villain's calling range. None of this is what you'd come up with from intuition; you have to play through it.

3. How variance feels at 15 BB

You'll Nash-correctly shove AK and lose to nines. You'll Nash-correctly fold KJ off, then see Maya's ten-six. Nash is the right strategy over a sample, not over a single hand. Playing 50 games and watching some of them go sideways teaches you that "I played that correctly" and "I won that hand" are different statements. That's a valuable thing to feel before you experience it for real money on bubble night.

What it's not

Maya is push/fold only. She doesn't make post-flop decisions because post-flop is where the math gets exponentially harder. A real solver computes mixed strategies across thousands of decision nodes; building a usable version of that in browser JavaScript is a different category of project. So she's a specialist, not a generalist. If you want to drill 50 BB cash game spots, this is the wrong tool. If you want to drill the spots that decide tournaments, it's the right one.

She also doesn't escalate blinds. A real sit-n-go would step up the blinds every 5 minutes; this one stays fixed at 50/100 to keep you in the push/fold zone. Once you've played through a session, you can mentally extrapolate: tighter ranges at 20 BB, looser ranges at 5 BB — same chart, different cutoff.

How to actually study with it

A few patterns that work better than just clicking buttons:

1. Pause before every shove decision. Even if you "know" the answer. Articulate it: "I'm SB with K9o at 14 BB effective, that's a fold per the chart." Then check the chart after the hand. If your articulated reason matched Nash, you're studying. If you said "K9 looks decent, shove," you're just clicking.
2. Note hands that surprised you. When Maya shoves something weirder than expected, look it up. She might've gotten unlucky on her randomization, or you might be misreading her stack depth. Either way, the surprise is the data.
3. Play 50 hands, then check your win rate. If you're below 50% you're either losing close decisions or making clear mistakes. If you're above 60%, Maya's variance is going your way and you should keep playing without getting cocky.
4. When you're stuck, pull up the flash-card trainer. Same Nash data, no game flow — pure 1-spot-at-a-time recognition drill with score tracking. Use the trainer to build the lookup, use Maya to apply it.

The bigger picture

Dexas Holdem started as a tournament timer for our weekly home game. Then a blind structure calculator, then a payout calculator, then an equity tool, then a Nash chart, then this — a real game you can play. Each piece individually is a 2-3 hour build; together they're a free toolkit that does most of what serious tournament study software does, except it runs in your browser, costs nothing, and doesn't try to upsell you to a $300 GTO solver subscription.

The trade-off is breadth-not-depth. Maya is a specialist; our equity calculator is a specialist; ICM is a specialist; bankroll is a specialist. None of them try to be everything. But for a home game player who wants to be measurably better next month — not "study GTO theory" better, just "stop making the obvious mistakes" better — together they're enough.

Maya was the last big piece. Everything else was math; she's the math made playable. Go beat her at her own game.

—

Want the rest of the home game playbook — blind structures, payouts, multi-table balancing, tournament etiquette? Grab the free 7-page Playbook PDF or check out the full tools index.