How a hobby computer vision project became the AI tool the world's top natural powerlifting team used to bring a world champion back from injury.
This case study didn't start with a client reaching out because their software was broken. It started with our team messing around with AI for fun.
2021. Before ChatGPT. Before anyone was talking about AI agents. Jon was doing his Master's in AI and bored with the standard coursework. So he combined two things he cared about — fitness and artificial intelligence.
He started training computer vision models to analyze his own lifts. You film a squat, the AI breaks down your body mechanics — frame by frame, joint by joint. Technically complex, the kind of project most people wouldn't attempt. He shared it on social media.
If you don't know them — they're the top natural powerlifting coaching team in the world. Very analytical. Very scientific. They saw what Jon was building and started using it to train their top athletes.
Their coaches did biomechanical video analysis for every athlete. Watch a lift. Frame by frame. Manually mark each body part — shoulders, hips, knees, ankles — tracking how they move through the entire rep.
One video took about an hour of skilled work. And these are coaches who charge premium rates.
The best biomechanical analysis methods require data you simply can't collect manually. A human can eyeball joint angles, but they can't precisely measure how much each muscle group contributed to a lift. That level of detail was off the table with a manual workflow.
The concept was straightforward. Take what a human coach does in an hour and have AI do it in a minute.
The execution was anything but simple. This wasn't one model doing one thing. Jon trained and fine-tuned multiple computer vision models, each handling a different part of the analysis. The real work was making them work together — feeding the output of one into the next, so you get a complete biomechanical breakdown from a single video upload.
The output isn't a vibe check. It's numbers. Quads: 45%. Glutes: 30%. Lower back: 25%. The kind of data that lets a coach make a real decision instead of a guess.
When you're healthy, your body distributes load in a specific pattern. After an injury, something shifts — the injured muscle stops pulling its weight, and the others compensate.
The AI runs the analysis and shows it. Glutes 20% less involved than they should be. Quads picking up the slack. Now the coach knows exactly what to do. Isolated glute work for a week. The athlete films the same lift the next week. The AI runs it again. Glutes are back up. The balance is shifting.
That's exactly what happened with one of their top athletes — a world champion powerlifter. He got injured. They used the AI to track his recovery session by session. They could see which muscles were falling behind and which were compensating. They adjusted his training based on that data every single week.
He came back. And broke his own squat record.
Since then, multiple universities have started working with us to benchmark the models against lab-grade equipment. The accuracy is holding up.
We've shipped agentic AI products since early 2025. If the standard tools don't solve it, we go deeper.