Inspiration
Poker rewards more than strong cards. It rewards understanding opponents, adapting under pressure, and knowing when aggression stops working. We built Jester Bot to explore whether an AI could combine an unpredictable playing style with disciplined mathematical decisions.
The TAMU Datathon’s heads-up Short Deck Omaha challenge gave us the perfect arena: incomplete information, unfamiliar hand rankings, and a strict computation budget.
What it does
Jester plays heads-up, five-card Short Deck Pot-Limit Omaha autonomously. It estimates hand strength, tracks opponent behavior, and adapts its strategy throughout a match.
Its four strategy gears give it different personalities:
- MANIAC: opens with frequent, randomized aggression.
- AGGRO: combines strong value bets with selective bluffs.
- VALUE: reduces bluffing against opponents who call too often.
- TRAP: plays strong hands more patiently against aggressive re-raisers.
Jester also detects repeated check-fold behavior and tracks passive timing signals behind a conservative reliability gate. Legal-action validation and safe fallbacks help keep it playing when something unexpected happens.
How we built it
We built Jester in Python, using FastAPI and Uvicorn to communicate with the tournament judge.
Preflop decisions use a lightweight heuristic based on pairs, suitedness, connectivity, and high cards. Postflop, Monte Carlo simulations sample possible opponent hands and remaining board cards to estimate equity. Jester compares that estimate with pot odds to guide its decisions.
An opponent model tracks folds, calls, raises, re-raises, and showdown information. A gear selector uses those observations and recent results to adjust the strategy, with minimum sample requirements and hysteresis to avoid switching too frequently.
We also built a testing league with scripted opponents, immutable champion snapshots, reproducible seeds, and automated match reports. Duplicate matches swap seats while preserving the deals, helping separate strategy quality from card luck.
Challenges we ran into
The biggest challenge was balancing decision quality with speed. More simulations improve equity estimates, but every decision draws from a limited time bank. We used fixed sample budgets and a global time governor to control computation.
The game’s rules also demanded care. Omaha requires exactly two hole cards and three board cards, and Short Deck changes familiar hand rankings. A seemingly powerful hand can be much weaker when evaluated correctly.
Timing signals proved especially difficult. Judge overhead and action sequences can resemble opponent tells. We rejected experimental changes that failed our safety or performance checks, preserving the validated champion instead.
Accomplishments that we're proud of
Our final fast test suite passed 296 tests, covering strategy behavior, legal actions, state resets, fault handling, and evaluation infrastructure.
In a completed internal benchmark against 15 practice opponents across 20 seeds each, Jester earned 262 out of 300 match points, averaging approximately +2.85 big blinds per deal, with zero bot errors or forfeits.
We also expanded our evaluation infrastructure to 62 training opponents, 12 reserved holdouts, and five test-only oracle opponents. Full validation of that expanded field remains ongoing.
What we learned
Aggression works best when it adapts. A strategy that overwhelms a passive opponent can become expensive against someone who calls selectively or re-raises well.
We also learned that reliable evaluation is part of building the bot itself. Reproducible experiments, confidence intervals, holdout opponents, and regression tests helped us distinguish promising ideas from noisy results.
Most importantly, a clever strategy is only useful if it finishes the match safely.
What's next for Jester Bot
Next, we want to complete the expanded league evaluation, improve gear switching against selective callers, and explore opponent-range-weighted equity estimates.
We also plan to investigate position-aware aggression and more deliberate bet sizing. Each improvement will face the same standard: measurable gains against diverse opponents without sacrificing reliability.