Poker Bot with Computer Vision & Neural Networks
Figure 1: Custom HUD overlay displaying player statistics and game state extracted from screenshots
Overview
Built automated poker bot using computer vision, OCR, neural networks, and custom HUD to play on PokerStars.
Challenge
PokerStars encrypted all game data in memory. Chat logs, card data, user credentials all encrypted/inaccessible. Only username/password readable in memory. Professional players used paid HUD applications with historical stats. Needed reliable card recognition. Avoid bot detection. Create profitable algorithm.
Solution & Approach
Screenshot-based approach: capture screen every few seconds, extract ALL game state from visual data only. Built custom database to track all player actions. Calculated behavioral statistics. Implemented Monte Carlo algorithm for hand strength calculation. Custom HUD overlay on PokerStars window. 100% accurate card recognition using pixel-perfect detection with 4-color deck. Randomized mouse movements and timing. Started with random actions, evolved to position-based strategy, experimented with neural networks.
Outcome & Impact
Profitable on 1€ cash tables. Played 4 tables simultaneously. Built complete player database and analytics. 100% accurate card recognition. Unprofitable on 5€+ tables (stronger competition). ~10€/week profit not commercially viable. Neural networks underperformed custom strategy (2014 technology limitations). Poker market declining. First experience with neural networks, embeddings, and training data preparation.
Technologies Used
Key Highlights
- ✓Screenshot-based state extraction (all data encrypted)
- ✓Custom HUD with player statistics database
- ✓100% accurate card recognition using pixel analysis
- ✓Profitable on low stakes, played 4 tables simultaneously
- ✓First experience with neural networks and embeddings