Their First Startup Got Catfished by Fake Athletes. Their Second One Fixed It With a Phone Camera.

Cordelia King's first company, TrainStop, connected local sporting clubs with players looking for a new team — press coverage at the time called it "Tinder for local sport." The comparison turned out to be more accurate than anyone intended. "People were saying that they were better than they were," King says, "and sometimes clubs were getting catfished."

The obvious fix was to make players add real stats to their profiles. The problem was that real stats for local-level sport barely existed. Below the professional leagues, the only numbers anyone tracked were goals scored or runs made — useless if you played defence, and inconsistent across leagues that didn't record things the same way. "A lot of local sport is filmed now," King says. "That footage existed, and the data could be attainable. It was just how to get it in a way that was scalable."

The oval that broke the model

The natural place to start was AFL — Australian rules football — because TrainStop already had distribution into hundreds of clubs desperate for exactly this data. It was also, it turned out, close to the hardest possible sport to build computer vision for. The oval is enormous, forcing constant zooming that makes it hard to track position. Eighteen players a side means constant occlusion — players running in front of each other, disappearing and reappearing in frame. And the statistics themselves are oddly shaped: goals and hitouts are relatively easy to detect, but clearances and handballs look different every time. "We were kind of like Atlas," King says, "pushing this rock up a hill."

The breakthrough came from a co-founder's hunch, not a data insight. Sam, who'd built his first basketball shot-tracker at fourteen, suggested switching sports entirely. "The court is the same size every time," King recalls him arguing. "It's five players, you can see everyone, the stats are really easy." The AFL model — already fine-tuned for occlusion and court-line detection through months of failed effort — worked on basketball almost immediately. The dead end hadn't been wasted; it had trained the exact skills the pivot needed.

"The court is the same size every time. It's five players, you can see everyone, the stats are really easy."

The unglamorous part of computer vision

What actually turned out to be hard, King says, isn't the flashy demo. Plenty of people post ten-second computer-vision highlight reels on LinkedIn built on off-the-shelf models. Processing ninety minutes of real footage — including timeouts, fouls, and dead time where the model has to learn that nothing worth analysing is happening — at usable cloud-compute cost is the genuinely unsolved problem. The eventual goal, she says, is running the model on a user's own phone rather than in the cloud at all — the same shift Strava made for running — which would push compute cost to zero and let them give stats away free to every player.

What it cost to build it

None of the three co-founders took a salary for a long stretch. Kai sold his car. Sam dropped out of university. King moved back in with her parents, leaving a stable graphic design job in fashion. "My mum wasn't happy," she says simply. Two thousand games later, and with the team now relocating from Melbourne to Austin, Texas — more than half of Superstat's users are already American, despite what King describes as almost no active marketing there — she doesn't second-guess the decision, only the timing. "I would definitely say not starting earlier," is her only real regret.

What's next: the hardest occlusion problem in sport

American football is next on the list, and King is clear-eyed about why it's difficult: a play can involve twenty players converging on one point, producing occlusion far worse than anything AFL threw at them. But the sport also has structural advantages basketball didn't — elevated stadium bleachers give a clean, consistent filming vantage point, every player has a visible number, and the field's marked lines make distance and position easy to calculate. The size of the prize is what's pulling them toward it: college football recruiting can determine a player's shot at contracts worth millions, and King says the recruiting process underneath it all is still almost entirely manual, with real performance data close to nonexistent below the very top level — precisely the gap TrainStop first ran into with local football, just with far more money riding on the outcome.