Hobbyist Builds Tennis Coach With Roboflow. Domain Constraint Does The Heavy Lifting. Obviously.
A builder combined Roboflow Agent, Claude Code, and a fine-tuned RF-DETR model to train a computer vision tennis coach. The pipeline works because tennis, like the parallel dance-sync project and prior basketball and golf swing trackers, benefits enormously from domain constraints that reduce what would otherwise require massive curation in messier visual domains.
This illustrates a principle I like to call constrained domain advantage. When your subject matter has rigid physical rules, a ball travels in a predictable arc, a swing follows a repeatable plane, your model needs far less data to reach competence. The mechanism is simple: narrow worlds produce clean signal. Messy worlds produce noise. Stop trying to build general vision systems first. Pick a domain where the physics does half your work for you.
An AI expert named Yash, who claims over 300K learners, is running workshops around these builds. Roboflow ML engineer Piotr Skalski separately published a VLM benchmark showing GPT-5.6 Sol represents a significant leap for OpenAI on object detection.
- Go to Roboflow, create a free account, and upload 50 to 100 images of yourself or someone performing a tennis serve. Label the ball, racket, and key body joints in each image. You now have a annotated dataset, which is the boring prerequisite everyone skips and then wonders why their model fails.
- Train a base model using RF-DETR or YOLOv8 on those labeled images directly inside Roboflow's browser interface. Export the trained model weights when training completes.
- Ask Claude Code to write a Python script that loads your model, feeds it a video of your serve, and draws bounding boxes plus a timing metric for contact point. The expected outcome is a video overlay showing where detection worked and, more importantly, where it did not, which tells you what to label next.