Clutch

Clutch, a Copenhagen startup, puts an AI camera on padel courts and turns every match into highlights and stats, in more than 150 clubs.

I was Clutch's first high school hire. I did three stints on the machine learning side and shipped 45,700 lines of code to production for more than 40,000 players.

My first job was finding the court. From the back fence, glass, mesh and players on the lines occlude its corners.

My early versions only found the corners they could see, so one player on a line scuppered the corner.

I modified TrackNet to predict six keypoints as heatmaps, and when one is hidden, a homography from the rest puts it back. 98.46% land within 15 pixels, and the model runs on CPU through ONNX on Fly.io. It undergirds some 150 million highlight views.

Those stats come from the match-analysis pipeline, which I extended from singles to doubles as one Docker-based inference system.

Highlights play on a TV at the club.

I also prototyped a coaching studio: I mined 402 clips from 42 YouTube coaching videos and searched them with RAG, so a described shot returns the right drill, and I presented it to the Clutch team. The hard part was relevance. Early answers drifted from the clip on screen, so I tied each tip to one clip's topic.

My first term was prototypes. By the third, my new work went almost straight to production.