We introduce an innovative application of computer vision and artificial intelligence to analyze training videos of canoe athletes preparing for the Olympic Games. We show the bottom-up approach of applying AI on real world problems and give an in-depth view on pitfalls and lessons learned. Our method employs foreground-background separation for canoe detection and waterline derivation. Through pose detection, we identify the paddle and have trained a neural network to recognize essential paddle positions for routine training analysis. Additionally, we incorporate biomechanical insights in a post-processing step to refine AI results and enhance analysis accuracy. Traditionally, biomechanics engineers manually screen training videos frame by frame to locate specific paddle positions and measure the paddle's angle relative to the waterline; a process taking about 20 minutes per athlete. Our approach significantly streamlines this process, reducing the workload by an order of magnitude.