Technical Communications · Solutions Architect Interview

Two Weeks to Kickoff

A fake soccer ball, a live broadcast to millions, and why we built a controller instead of retraining the model

Yariv Barsheshat · September 10, 2026
Case: real-time computer vision for virtual ad replacement — mtl.ai, Montréal · UEFA Euro 2024

yariv@barsheshat.com1 / 5

The business problem

A billboard ad started fooling our model,
14 days before Euro 2024

The situation

Virtual ad replacement: camera feed → our models find the billboards and everything in front of them → the ad is swapped digitally → to air, live
A new stadium ad featuring a fake soccer ball appeared. The production model detected it as a real ball — a visible artifact on air every time the ad rotated in
Stakes: millions of viewers, sponsors paying for clean replacement, no retake in live sport
Constraints: 14 days, no downtime, and the fix could not degrade the 99% of frames that were already right

Business needs

Zero visible artifacts on air, from match one
No regression anywhere else in the broadcast
Operators at the stadium can act without an engineer on call
Ready inside two weeks, with a way to verify it

Technical needs

Detections by class (ball vs players) instead of one merged class
Thresholds that vary by region of the frame and by time
A control surface operators already understand
Validation that doesn't require a full retrain cycle

Two Weeks to Kickoff2 / 5

Options and tradeoffs

Three options, honestly weighed

Option	Time to safe	What you gain	What it costs	Verdict
Retrain the model with the new ad in the data	Weeks: collect footage, train, full validation pass	Fixes the root cause inside the model	No guarantee performance stays consistent elsewhere; a new model to re-prove; nothing to roll back to if it regresses on air	Not in 14 days
Raise the ball threshold globally	Hours	Trivial to ship	Real balls missed across the whole pitch — the one object the system must track	Wrong trade
Regional, class-aware threshold controller	Days; the model itself is unchanged in weights	Surgical: only that class, only that region, only while the ad shows. Reversible in one control. Testable in isolation	Manual: relies on operator attention and a priori knowledge of where the ad appears. A fix around the model, not in it	Chosen

The tradeoff in one line: an operational fix over a model fix — certainty by kickoff, paid for with a human in the loop.

Two Weeks to Kickoff3 / 5

The architecture

Two changes: one output layer, one controller — no retrain

Broadcast camera feedlive frames from the stadium

→

Detection modelchanged: output layer now reports per class (ball · players · …) instead of one merged class. Weights untouched

→

Threshold controllernew: threshold = f(class, region of frame). Predefined regions · live adjustment · applied before detections reach compositing

→

Broadcast systemcompositing + ad replacement → to air

The model's weights never changed. Everything that already worked kept working; only the ball class, in chosen regions, got stricter on demand.

Operator control surfaceMPC-style pad controller the operators already use at the stadium

T − 60 mindefine regions where the problem ad is known to appear

T − 2 minad rotates in → raise ball threshold in that region with one control; lower it when it rotates out

Two Weeks to Kickoff4 / 5

Impact, limits, lessons

Zero false positives across the tournament

0

false positives triggered by the ad, across every Euro 2024 match

< 14 days

from discovery to running live at kickoff — [actual build days]

Unchanged

detection performance everywhere else — model weights never touched

What made it work

Talked to the operators first. Learned their controls and their reality — a 2-minute lead time — and designed to it
Chose the reversible option under a hard deadline
Separated the model's knowledge from the broadcast's policy

Limitations, and what's next

Manual, and relies on knowing where the ad will show
A fix around the model, not in it
Next: auto-propose regions from the ad schedule; log operator tunings as labels for the next retrain; keep per-class output for good

Two Weeks to Kickoff · Yariv Barsheshat5 / 5