Inspiration
We wanted to build a poker bot that makes sensible decisions under uncertainty while staying simple enough to trust. In a tournament where repeated errors cause a forfeit, reliability matters as much as strategy.
What it does
SteadyAce plays heads-up short-deck five-card Pot-Limit Omaha. It chooses among folding, checking/calling, and half-pot or pot raises. It estimates winning chances, compares them with the call price, and adapts its opponent range using observed aggression. Sudden all-ins receive an explicit equity-versus-price decision.
How we built it
We implemented the three strategy functions in the supplied Python/FastAPI agent. A NumPy helper evaluates simulated hands using exactly two hole cards and three community cards, including the variant's unusual hand rankings. Simulations reuse results when visible cards have not changed. Match-specific memory resets every half. The bot runs offline without API keys or a trained model.
Challenges we ran into
The main challenges were the eight-second thinking budget, correct Omaha hand construction, and handling aggressive opponents without treating every raise as either a bluff or an unbeatable hand. We tested all-in call prices, tied pots, capped raises, and half resets while preserving the organizer's server plumbing and rules.
Accomplishments that we're proud of
The submitted version passed 31 tests and won all 21 local duplicate matches across seven practice policies and three seeds: 42,000 hands with zero errors, forfeits, or exhausted time banks. Both native Linux and Linux AMD64 Docker build/smoke checks passed. These are practice results, not tournament results or a guarantee of winning.
What we learned
Pot odds and expected value provide a useful foundation for decisions under uncertainty. Reproducible match results and compatibility checks are stronger evidence than plausible strategy alone. Our policy uses game-theory-informed heuristics; it is not a solved GTO strategy.
What's next
Evaluate against more independent opponents and improve opponent-range estimates while preserving the small runtime and reliable action handling.
AI assistance disclosure
OpenAI Codex assisted with implementation, tests, documentation, and local validation. The playing bot makes no calls to an LLM or external service. See the repository's pokerbot/AI_DISCLOSURE.md for details. Include this disclosure in the project story or the event's AI-use field.
Links and submitted version
- Repository: https://github.com/RedRojoaka/datathon_2026
- Poker project and opening instructions: https://github.com/RedRojoaka/datathon_2026/tree/main/pokerbot
- Submitted bot commit: https://github.com/RedRojoaka/datathon_2026/commit/b3515aade1a437694345a24b93e1d9b5ee643abc
- Strategy at the submitted commit: https://github.com/RedRojoaka/datathon_2026/blob/b3515aade1a437694345a24b93e1d9b5ee643abc/pokerbot/STRATEGY.md
- Validation at the submitted commit: https://github.com/RedRojoaka/datathon_2026/blob/b3515aade1a437694345a24b93e1d9b5ee643abc/pokerbot/VALIDATION.md
