Describe what an embedding is and one product feature it makes possible.

Thrumcap Sports Analytics builds MatchPrint, a platform that turns match footage and tracking data into embeddings, numbers that place similar plays and similar players close together in the same space. Esk Valley FC, a club in the Northshire Championship, wants two things built on top of it: a tool that finds passages of play like the one an upcoming opponent just ran, and a tool that surfaces players who play like a transfer target. Rozenn Wrycliffe, Thrumcap's Head of Product, has one engineering pod for the quarter, and a board meeting on Friday where Esk Valley's head of recruitment, Tobias Sandoway, already knows which one he wants first.

How to answer this, stage by stage

Nobody is grading whether you can define a vector. They're grading whether you can name a real feature it unlocks, then defend which one you'd build first and which one waits.

Let's learn

What do you actually get when a computer turns a football match into a list of numbers?

MatchPrint is Thrumcap Sports Analytics' platform for professional football clubs. It reads a club's match footage and tracking data, and turns passages of play into embeddings: a list of numbers built so that plays with the same shape land close together in that list, no matter which players ran it or what the score was. It does the same for players across a season, turning a player's passing, carrying, and defending into a second list of numbers, so players who actually play alike sit near each other, even when their headline stats don't match at all.

Different matches, different minutes, different scorelines. The plays that share a shape still land in the same corner.

Before MatchPrint, Tobias Sandoway, Esk Valley FC's head of recruitment, ranked forty scouting reports by hand every week, from memory and from tape he'd watched three or four times each. He was good at it. He could watch four minutes of a winger's footage and tell you, before the ad break, whether the end product would hold up two divisions higher.

None of this was slow because Tobias was bad at it. It was slow because every judgment lived in his own head.

Now he has MatchPrint, and he wants two things built on it before the transfer window closes: Similar-play search, where an analyst drops in a clip of an upcoming opponent and gets back Esk Valley's own matches with the same shape of attack, and a comparable-player recommender, which would sit inside the club's own dossier template and auto-suggest transfer targets who play like a name the committee already likes.

Here's the turn. Both features run on the same kind of embedding. Only one of them is actually ready.

Similar-play search draws only on Esk Valley's own match archive: every game the club has played is filmed by the same broadcast feed, tagged the same way, complete from day one. The comparable-player recommender needs something bigger: enough clean, representative player data across every league MatchPrint touches, so a player's numbers actually mean the same thing wherever he plays. Right now, only 5 of Thrumcap's 14 leagues clear that bar. Esk Valley's own league, the Northshire Championship, isn't one of them yet. A data-provider swap left six weeks of matches half tagged partway through last season.

What it costs at its worst: ship the recommender now, and it starts surfacing "comparable" transfer targets straight into a boardroom dossier, built on the thinnest data MatchPrint has. Get one of those wrong in front of the committee, and Esk Valley doesn't just stop trusting the recommender. They stop trusting Similar-play search too, the one tool that was actually right.

What I would leave alone: Similar-play search doesn't need any of this. It runs on a club's own footage, which is always complete the day a club signs up, so there's no coverage bar to clear before it ships to anyone.

The lesson: a feature isn't safe just because the model runs without erroring. It's only safe where the data behind it is actually good enough to trust, and that's a different map for every embedding feature you build, not one map for the whole platform.

Now here is the same thing as a story

The short version above is what you actually say in the room. Read this one for the two years it took a promise made to two leagues to catch up with a Tuesday in Northshire.

Tobias Sandoway can watch four minutes of raw tape and tell you, before the first commercial break, whether a winger's end product will hold up two divisions higher. He's run recruitment at Esk Valley FC for six years, on a laptop with a cracked hinge and a spreadsheet nobody else can read. Every Sunday night he ranks forty scouting reports by hand, in the order he trusts them, not the order they arrived.

MatchPrint arrived in the spring. For the first few months it did the small, boring thing well: an analyst could paste in a clip from an upcoming opponent and get back every match in Esk Valley's own archive with the same shape of attack, the same run down the same channel, finished the same way. Nobody's job changed. It just meant an assistant coach spent Tuesday morning watching three relevant clips instead of trawling six matches for one.

Then, at a demo in June, Rozenn Wrycliffe showed Tobias something bigger: a version of the same idea, but for players instead of plays. Type in a transfer target's name, and it would surface players from anywhere in the world who actually played like him, not just players who matched on the usual filters like age and goals per ninety.

Tobias didn't sleep well that week. Esk Valley's board meets every Friday to review transfer targets, and he'd been arguing for a winger from three leagues over that nobody else in the room had heard of. A tool that could say, in one screen, "here are four players who move like him," would have made that argument for him.

So on the Tuesday before Friday's meeting, he asked Rozenn to ship it. Not the play-search tool, that was already live and useful. He meant the player one, wired straight into the dossier template the board actually reads.

Three days between the demo and the meeting. Nowhere near enough time for the real question to get asked.

Rozenn said no, and had to explain why on a call with four minutes left before her next meeting.

Here's what she told him. MatchPrint's player embeddings are only as good as the event data underneath them, every pass, every carry, every defensive action, tagged consistently enough that a player's numbers mean the same thing in Northshire as they do anywhere else. Right now, that's true in five of Thrumcap's fourteen leagues. Esk Valley's own league isn't one of them. A data provider swapped systems mid-season and left six weeks of matches half tagged.

Rozenn's team had already run the cheap check, the one that should always come before a feature like this ships. Take thirty pairs of players a scout had already judged genuinely comparable by hand, drawn from leagues that vary in coverage, and see how often the model's own top five matches actually included the real one.

In the two founding leagues with full coverage, it hit 24 of 30, eighty percent. In leagues like Northshire, it hit 15 of 30. Half. A coin flip, and Tobias wanted to put a coin flip in front of his board on Friday.

Tobias didn't take it well, not at first. He'd seen the demo. It worked in the room. What he hadn't seen was that the demo ran on Thrumcap's own two founding leagues, the two with a full year of clean data behind them, not on Northshire.

Here's the decision I'd take back, and it isn't Tobias asking for the feature early. Given what he'd seen, that was a completely reasonable ask. It's a decision from eighteen months earlier, when MatchPrint only served those first two leagues. Thrumcap's own onboarding deck told every new club that comparable-player search worked the same everywhere, out of the box, because at the time it did. Nobody built a visible flag for what happens once the platform outgrows that promise.

Run Friday's meeting again, with that flag already built. Tobias still asks Rozenn for the player tool on Tuesday. This time, MatchPrint tells him plainly: Northshire isn't covered yet, here's the honest date it will be, about two quarters out, once coverage crosses the bar the same nearest-neighbor test checks. Friday's board meeting runs on Similar-play search and Tobias's own judgment, the same as it always has. Nothing breaks. Nothing goes into a dossier that hasn't earned its place there yet.

The two designs, side by side: one hands every club the same feature list on day one and hopes the data catches up quietly. The other tells a club exactly what's ready and what isn't, and lets that honesty do the selling instead of the demo.

What I'd tell myself, watching that first onboarding deck get written: a promise that's true for two leagues doesn't stay true once you're proud enough to sell it to twelve more. Someone should have asked, the day league three signed up, whether the promise still held. Nobody did, until a Tuesday two years later made it impossible not to.

ORDER, for picking which embedding feature earns the first quarter

PICK would fit if this were only two options with an even trade. But Esk Valley wants both, eventually, and the real question is which one goes first without breaking the other. That's ORDER's job.

The third box is the one Thrumcap is still waiting on, eighteen months after the second box became true for the first two leagues.

One of these you fix by ignoring a bad result. The other one is already sitting in a meeting minute by the time anyone notices.

Three things worth stating directly, since the real judgment sits here. The alternative worth naming and rejecting is auto-clustering raw footage into highlight reels first, since it demos well and needs less new infrastructure. It loses because it never touches a real decision Esk Valley's staff make, so shipping it would prove nothing about whether MatchPrint's embeddings are actually any good. The AI-specific failure worth naming is silent coverage drift: a model that runs, without erroring, in every league, even the ones where the underlying event data is too thin to trust, because nothing in the pipeline stops it from returning an answer. The guardrail is the same nearest-neighbor check, run per league, with results gated below the bar instead of a confidence label nobody reads under deadline pressure. And the trade-off is accepted on purpose: holding the recommender back costs Thrumcap a flashier board pitch this quarter, and costs Tobias a tool he wanted by Friday. That's slower, on purpose, in exchange for never putting a coin flip in front of a transfer committee.

And if you want to be sure it really works, try it somewhere else

Same five letters, a hospital corridor instead of a boardroom, and this time the thing nobody separated is a healthy stray marking from a real tumor.

Bracken Hollow Veterinary Diagnostics runs ScanMatch, which turns a dog or cat's ultrasound or x-ray into the same kind of embedding: an image with a similar shape and shadow pattern lands close to other images like it, so a vet reviewing a hard case can pull up the hospital's own past scans that looked the same and see what they turned out to be. Bracken Hollow's lead radiologist, Ingrun Ondiek, wants a second feature built on the same embeddings: an automatic second-opinion flag, cross-referencing rare breed-condition pairs the hospital's own case history barely covers.

Different hospital, different image, the same missing coverage check.

Same chart, mapped onto Bracken Hollow. Outcome: protect a vet's ability to trust a comparison enough to act on it in a real appointment, not just "add AI to radiology." Reversibility: the same-hospital comparison tool is reversible, a vet just ignores a bad match and moves on; the automatic second-opinion flag, once it starts appearing on official scan reports, is not, a wrong flag on a client's record is hard to quietly take back. Dependency: the comparison tool depends only on the hospital's own scan archive, complete and consistent since Bracken Hollow only reads its own machine's output; the second-opinion flag depends on enough labeled cases of a specific breed-condition pair existing anywhere in the hospital's history, and for most rare pairs, that number is close to zero. Evidence: the same nearest-neighbor check, run on a sample of scans a senior radiologist already diagnosed, splits cleanly, common conditions score well past 80 percent, rare breed-condition pairs barely clear a third. Rank: build the comparison tool first, ship the second-opinion flag only for condition pairs with enough labeled history to clear the bar, and say plainly, on the report itself, when a pair doesn't have enough history yet instead of guessing quietly.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: ship the feature whose data you already fully own, hold the one that depends on data you don't have enough of yet, because a shared dependency doesn't care how excited anyone is to ship.
Cost: no budget to properly clean the thin-coverage leagues this quarter. Ship a visible coverage flag instead of the feature itself, so nobody mistakes "not ready" for "broken."
The model got better, for real: say MatchPrint's next embedding model is measurably stronger across the board. Keep the same coverage gate anyway. A better model can still be trained on the same thin data underneath one league, and better was never the same claim as ready.

Where people run it wrong.
They ship the feature that demos best, not the one whose dependency is actually satisfied.
They add a confidence label and call the coverage problem solved, when nobody under deadline pressure reads a label.
They test the nearest-neighbor check once, at launch, and never rerun it as new leagues or new case types get added.

How to use it live. Before ranking anything, ask out loud: "which of these features runs on data we already fully own, and which one needs data we're still collecting?" Whichever answer comes first is usually the order an interviewer is actually listening for.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0