I’m building a proof-of-concept that watches a live online-casino roulette table, recognises each winning number directly from the streamed video, and immediately places the next bet according to a colour-based Martingale sequence. The aim is not to print money but to test and refine betting ideas, so accuracy and repeatability matter more than aggressive staking. Here’s the flow I have in mind: • Computer-vision module (OpenCV, Tesseract, or another proven stack) captures the wheel’s history panel or number ticker, then converts it to a clean numeric result with at least 95 % reliability. • A lightweight ML layer (you may use a simple classifier or CNN you’ve trained) should handle variable fonts, dealer hands occasionally blocking the view, and different casino themes. • Decision logic: if the last result is red, double the stake on black (and vice-versa) until recovery, following classic Martingale rules. All parameters—base chip size, stop-loss, number of steps—must be editable in a config file. • Betting execution must interact with the browser DOM or the game’s HTML5 canvas directly; no hard-coded autoclicker coordinates. Selenium, Playwright, or a custom JavaScript injection are all acceptable so long as the solution adapts to table resizing. Deliverables 1. Well-commented source code (Python preferred, but I’m open). 2. A short model-training notebook or script showing how the number-recognition network was produced. 3. Setup guide and a video or live call demonstration proving the bot can (a) read live spins with ≥95 % accuracy, and (b) place ten consecutive bets on the correct colour without manual input. If you’ve previously handled OCR or computer-vision projects in dynamic, low-light streams, you’ll likely breeze through this. Let me know which libraries you’d lean on and any clever tricks you’ve used to stabilise recognition under flickering casino lighting.