Odomo — The World's First AI for the 28 Card Game
28 (also known as Irupathiyettu, Twenty-eight, or Kerala 28) is a fast, deeply strategic trick-taking game played across South Asia. Odomo brings it online — with the first reinforcement-learning AI ever trained to bid, play, and reveal trump like a seasoned 28 player.
What is the 28 Card Game?
28 is a four-player, two-team trick-taking card game popular in Kerala and across the South Indian diaspora. Players are dealt cards in two stages, bid for the right to set a hidden trump suit, then play eight tricks attempting to capture 28 total points from the deck — concentrated in just four ranks per suit (J = 3, 9 = 2, A = 1, 10 = 1).
For a full walkthrough of bidding, tricks, hidden trump, and scoring tiers, see How to Play.
Why an AI for 28 is hard
- Imperfect information. Trump is hidden until revealed. Opponent hands are private. Unlike chess or go, optimal play depends on probabilistic reasoning over what you can't see.
- Two-stage bidding under uncertainty. You bid first on a 4-card hand, then again on an 8-card hand. The right bid depends on your partner's likely holdings — which you can't see.
- Partner coordination without communication. 28 is a partnership game with no signalling allowed. Cooperation has to emerge from shared strategy and observed play.
- Strategic trump reveal timing. Reveal too early and you give away information; too late and you lose tricks you could have won. Choosing when to break trump is the strategic heart of 28.
How Odomo's AI was built
Odomo's AI is a deep reinforcement-learning agent trained with Proximal Policy Optimization (PPO) through self-play. Over 100,000+ episodes, four copies of the agent play against each other, refining a policy network that maps an observed game state (own hand, public history, bidding context, trump-reveal state) to a legal action distribution.
Training runs in TensorFlow on Apple Silicon via the Metal backend. The trained model is deployed as a drop-in seat in our multiplayer server — and the game is engineered with a rule-based fallback so even if the model is offline, the table keeps playing.
Three ways to play
vs Bots
One human, three rule-based bots. Instant pickup, no signup. Great for learning the rules or a quick game.
vs ML AI
Challenge the trained model — three Odomo agents seated as your opponents and partner. Sign-in required.
PvP — 4 humans
Real-time over sockets, with persistent rooms so anyone who disconnects can reconnect to the same hand.
Frequently asked questions
- Is the 28 card game free to play on Odomo?
- Yes. You can play 28 against rule-based bots without an account. Multiplayer (PvP) and challenges against the trained ML AI are also free — they only require a Google sign-in so we can save your seat across reconnects.
- How was the 28 card game AI trained?
- Odomo's AI is a deep reinforcement-learning agent trained with PPO (Proximal Policy Optimization) over 100,000+ self-play episodes. The observation space includes the agent's hand, public play history, bidding context, and trump-reveal state. Training runs on TensorFlow with Apple Metal acceleration.
- What makes 28 hard for an AI?
- Three things: imperfect information (hidden trump and unseen opponent hands), two-stage bidding under uncertainty, and partner coordination without explicit communication. The strategic decision of when to reveal trump is the heart of the game.
- Where can I learn the rules of 28?
- See our How to Play page for a full walkthrough of bidding rounds, tricks, hidden trump, and scoring tiers.
- Is my game state saved if I disconnect?
- Yes. Active rooms are persisted server-side; you can rejoin any in-progress game from the home screen using the Continue button.