College basketball’s new assistant coach: AI

Open this photo in gallery:

The season was off to a rocky start for the men’s basketball team at the Virginia Military Institute. The Keydets, as they are known, had won six games but lost 12 overall as they headed into another matchup on a recent Saturday night in their hometown of Lexington. They were due to face The Bulldogs from The Citadel, a college in South Carolina. Both are Division I teams in the National Collegiate Athletic Association (NCAA). Since each represents a military school, the rivalry is a little more intense than the norm.

But the Keydets had an advantage that night: a system powered by artificial intelligence that had ingested and analyzed hours of game footage, practice footage, reports from the coaches, details on players and opponents, and had also learned the team’s custom terminology and playbook. A chatbot sat on top of this data repository so that coaches could ask the system, called PickAndRoll.AI, for pre-game reports, drill down into the statistical minutiae of action on the court, and get recommendations for plays to call during live matches.

Pick and Roll was created by a startup called The Intelligent Search Company (TISC), run by two unflappably enthusiastic 20-somethings. A dozen NCAA teams are testing the technology, but the game on Saturday marked only the second time one of them would be using it live.

That probably explained why co-founder and chief executive Arpan Bhattacharya’s knee was bouncing up and down as he sat on the sidelines with his laptop open before the game started. Around him, cadets in military fatigues piled into the stands and hurled insults at the Bulldog players trying to warm up on the court. To pass the time, Mr. Bhattacharya did some googling to find out the species of the Keydet’s mascot, Moe, who was bopping around to a brass band oom-pahing in the corner. (It turned out to be a kangaroo.)

Open this photo in gallery:

Shortly after tipoff, a Pick and Roll update appeared on his laptop: “Braxton Williams misses 27-foot three point jumper.” The message arose a few seconds after the Citadel player indeed missed a three-point shot. It was part of an AI-generated play-by-play, tracking every pass, shot, rebound and foul in the game. The company was vague about how it was producing the summary. ESPN was broadcasting the game and ran a seemingly AI-generated play-by-play online. It wasn’t clear if TISC was doing its own AI analysis from ESPN’s video feed, or simply accessing the broadcaster’s summary. (The company said it works with a variety of data providers for live and historical footage.)

In any event, the text summaries and live video footage were fed into Pick and Roll so that it was working with the latest information. That was the easy part. The real test was whether VMI’s coaches would ask for input during the game – and whether they would receive good advice. Mr. Bhattacharya flipped through tabs on his laptop and waited.

He and his co-founder, Moe Sabbaghi, didn’t intend to bring AI to basketball. The two met while working at e-commerce company Wish in San Francisco, and fixated on the idea of applying AI to high-stakes situations to surface crucial nuggets from data to help people make decisions. These scenarios, such as law enforcement, firefighting and military deployments, share a few characteristics, Mr. Sabbaghi explained. “There’s a lot of data, there’s limited time to make a decision, and there’s some kind of changing environment that makes decision-making difficult, or an adversary working to throw you off,” he said.

They started TISC in May, 2024, after Mr. Sabbaghi dropped out of the computer science program at the University of Waterloo, and raised US$2.1-million in pre-seed funding, including from N49P Partners in Toronto and Panache Ventures in Montreal. Mr. Bhattacharya lives in San Francisco, where TISC is headquartered. Mr. Sabbaghi, who was born in Iran and grew up in Mississauga, works in Toronto. TISC has five other employees today, with strong Canadian representation. “I’m a believer that talent is the greatest export Canada has to offer,” Mr. Sabbaghi said.

Open this photo in gallery:

Throwing AI at life-and-death situations is an ambitious gambit for a fledgling startup. Fortunately, sports proved to be a less fraught domain in which to build their technology. Basketball has more games per season than some other sports, and it didn’t hurt that the founders are fans. “I’m a Raptors fan, brother,” Mr. Sabbaghi said. “I had to become one, because everyone in school was non-stop talking about them.” That wasn’t too long ago. He’s 23.

The desire to expand to military applications some day makes VMI a fitting testbed for the technology. Founded in 1839 and entwined with American military history, it’s the kind of place that not only has a Civil War museum on campus, but one where the gift shop sells socks featuring Confederate general and one-time faculty member Stonewall Jackson.

Associate head coach Xavier Silas first heard about TISC through a contact with the NBA, and was immediately intrigued. “Whoever figures out how to use it to help win first is going to have an advantage,” he said. “For a while.” TISC’s goal, of course, is to sign on as many customers as possible, eventually eroding any edge it may provide to early adopters. For now, he’s figuring out what Pick and Roll can do.

Video analytics have been around for years, and there are a few products on the market to help teams study and catalogue digital footage. Montreal’s Sportlogiq, for example, was founded in 2015 to use computer vision and machine learning to analyze video footage and provide data and insights to hockey, football and soccer teams.

What’s different about Pick and Roll is the ability to ask questions and get detailed information much faster than in the past. Or as Mr. Silas put it: “Go down a rabbit hole.” Perhaps he wants to know how many times a certain player runs to the corner after switching from defence and offence. He can tap Pick and Roll for that info.

NFL’s use of AI to predict injuries aims to keep players healthier

In the past, he might have reviewed hours of game footage to get those numbers. He has plenty of experience with that gruelling ordeal from his work with NBA teams. For the Detroit Pistons, he once reviewed an entire season to calculate the significance of scoring from different positions near the basket. These projects meant keeping his eyes glued to the screen, watching footage in bed or at the pool with his kids, and trying not to miss anything to accurately record his findings. These insights can be useful. Last year at VMI, he discovered there was a 20-per-cent difference in the success of a three-pointer when preceded by a good pass versus a bad one. The team will shout “20 per cent!” as a reminder. With TISC, he can get more specific, asking about the success rates for each player in that scenario.

The mountain of information at his fingertips with Pick and Roll is a powerful thing, but making use of it is another matter. “I’m not going to sit here and tell you I have it all figured out on how to use it and how it affects winning,” Mr. Silas said. “But I know it’s coming.”

At Western Michigan University in Kalamazoo, Sam Little is starting to get a feel for the platform. As an assistant coach for the Broncos, the men’s basketball team, he has been testing Pick and Roll to see if the reports it produces are in line with what he and his colleagues are seeing. If not, he probes the system for data to back up its findings. “It confirms a lot of what we see, which is a good thing,” he said. “Then there are times where it points out something that just isn’t relevant to our style of play.”

The system might recommend that a certain player see more game time, but the coaches have other ideas. “There are various factors AI can’t account for, just from us being here every day and having relationships with our players,” he said.

Neither coach seemed to expect the system to produce an earth-shattering insight on its own. Generative AI models can process vastly more information than the human brain, but since they’re trained on existing human knowledge and function as big statistical machines, the odds of one producing a genuinely novel finding is slim.

University of Waterloo, Orioles partner on AI project to track pitcher biomechanics

Still, AI can be full of surprises. Google’s DeepMind division famously built an AI model to play Go, a strategy board game. Called AlphaGo, the model trounced champion player Lee Sedol in a tournament in 2016, in part by making a move so unexpected that it left Mr. Sedol puzzling over his response for close to 15 minutes. The unconventional play, known as move 37, turned the course of one game. Later, DeepMind built another version called AlphaZero that taught itself to play the game from scratch, without learning first from humans. The feat suggests that AI systems, free from human biases, can develop powerful strategic knowledge.

There are many more variables at work in a basketball game than in Go, but the equivalent breakthrough for TISC would be for its system to recommend a play during a game that the coaches did not consider – and it proving decisive. “We’re hoping for it to happen soon,“ Mr. Bhattacharya said.

Back at the Keydets game, problems were arising on and off the court. VMI’s players were trailing, while the cadets in the crowd, already red-faced from shouting at the Bulldogs, hollered at the refs for making bad calls. Mr. Bhattacharya, meanwhile, noticed that the live gameplay updates on Pick and Roll were a few seconds behind the action, which meant that AI-recommended play calls might not be working with the whole picture. “We’re really trying to get that number down so it’s a little more useful,” he said. (The company is considering custom hardware to get around broadcasting delays.)

Coaches can access TISC’s platform on a mobile device, and the company set it up with common, pre-written questions so they can press a button instead of wasting precious seconds typing. Mr. Bhattacharya, in addition to monitoring Pick and Roll on his laptop, scanned the VMI coaches across the court in search of a mobile device. Mr. Silas was pacing with a sheaf of paper in his hands. No luck there. But he soon spotted another coach gripping an iPad.

Less than 10 minutes into the game, he pumped his fists in the air. “They asked a question!” he said. Not long after, VMI swapped out one of their players – the exact swap that Pick and Roll had recommended. (This turned out to be a coincidence, but it showed the system had the right idea).

Cathal Kelly: AI officiating gives us a grim look into our future

There was another similarity later. Pick and Roll had recommended the team lay off challenging a particular Bulldog player, and Mr. Bhattacharya’s ears perked up when he heard VMI’s coaches yelling the same thing. His eyes flicked between his laptop and the court. “I’m thinking of some questions that could be helpful that they’re not asking,” he said. “I’m going to text them.”

During the second half of the game, as the margin for error narrowed, the questions tapered off. The Keydets still trailed the Bulldogs. The buzzer sounded and sealed their fate. They lost 82-68.

Afterward, Mr. Bhattacharya saw that assistant coach Nick Korta had queried Pick and Roll 19 times, asking for timeout strategies, play calls and player substitution requests. “Using TISC on the bench was like having another basketball mind whispering ideas,” he said later. “It would give us stats in real time that would make us go, ‘Huh, I hadn’t even thought about that.’” Even obvious tactics can be easy to forget in the moment, he said.

The system was most helpful during timeouts, when the action stopped. “The game moves so fast and the software hasn’t reached a point yet where it can be fully effective within the flow.”

Setting aside what TISC might be able to do during live games in the future, the platform’s ability to hoover up and analyze footage to provide detailed insights raises an issue. Both Mr. Silas and Mr. Little at Western Michigan say the time they spend manually reviewing video makes them better coaches. “Bill Belichick, Nick Saban, all of them hang their hats on the grind of film,” Mr. Silas said.

TISC potentially allows coaches to forgo the trouble by asking for what they want. What, then, becomes of their abilities as coaches? Do any skills atrophy? And if AI tools become commonplace, will the next generation of coaches have a harder time getting an intuitive sense of the game? Difficult, time-consuming tasks build competence and confidence, no matter the domain, whereas the temptation to offload cognitive work to AI can be irresistible, especially under pressure.

“Losing that feel could be a consequence of relying on AI too much,” Mr. Little said. “There’s no substitute for us watching film.” Just as writing things down strengthens his memory, reviewing footage deepens his sense of the game, he said.

Mr. Silas, meanwhile, doesn’t think his film hours are going to change, but he’ll get more information from TISC during that time. It’s up to him to figure out how to use those findings, combined with everything else he knows about basketball and the players on his team. “You still have to have a feel, and emotional intelligence,” he said.

On Saturday, as the crowds left, Mr. Bhattacharya talked with an employee from VMI. What was up with the live feed of the game? The lag made it difficult to deliver timely play calls. Could they access it any faster? The employee didn’t know, explaining Mr. Bhattacharya would have to talk to ESPN about that. He nodded and walked off. It was a problem for another day.

Open this photo in gallery: