Skip to content
AktwinsTrust. Transparency. Technology

Case study

A talent-pool score that has to be the same every time you run the clip

A solo founder wanted a sports talent platform: athletes record the drill for their sport, upload it, and the system scores the performance. He had already vibe-coded a prototype. The scores were neither consistent nor accurate. We took that as a prototype engagement and made the computer-vision scoring hold.

ClientA solo founder building a talent-pool platform for sport — record a drill, get a score scouts and athletes can trust.

Computer VisionSports

Computer VisionSportsPrototype

Tech stack

  • Python
  • PySide6
  • OpenCV
  • YOLO
  • RTMPose
  • ByteTrack
  • FastAPI
  • React Native
  • Expo
All case studies

The problem

A talent pool is only as good as the score behind it

The idea is clear. An athlete records the drill that belongs to their sport, uploads the clip, and the platform assesses it — the same way, every time. Without that, the pool is just video and opinion.

He had already built a first version with vibe coding. It looked like a product. It did not score like one. Run the same clip twice and the count moved. Close contacts were guessed. Accuracy was not good enough to put in front of talent or investors.

What we built

Computer vision with rules, not a prompt that hopes

We treated this as prototype development, not a rewrite of the pitch. Football juggling was the first drill we used to prove the scoring: see the ball and the feet, decide what counts, and only then produce the score.

  • 01

    Record and upload the drill

    • The talent records the drill and submits the clip into the system
    • Analysis runs on the host. The phone is for capture, not for the model

  • 02

    See what actually happened in the clip

    • YOLO for ball and player, RTMPose for feet, ByteTrack to keep identities across frames
    • Overlay so you can watch the same contacts the engine is scoring

  • 03

    Validate before you score

    • Each contact is valid, invalid or ambiguous, with a reason when it is not counted
    • The official count is valid touches only. Ambiguous is shown, not guessed into the score

  • 04

    A score you can run twice

    • Assessment metrics come from the validated events, then a deterministic score
    • Same clip, same rules, same result — that is what the earlier prototype could not do

What changed

The prototype finally scored like a product, not like a demo.

  • Consistent scoring

    • The same footage produces the same count and the same score
    • That is the bar for a talent pool. A moving number is not a ranking

  • Accurate enough to stand on

    • Contacts are decided from ball, feet and trajectory, not from a generated guess
    • Where the clip is unclear, the system flags it instead of inventing a touch

  • A founder who can show the idea working

    • He still has the talent-pool story. He now has a prototype that scores a drill the way he described it
    • The engines are built to take the next sport's drill without throwing the scoring logic away

What this unlocks

Vibe coding got a prototype on screen. It did not get a score you can defend. We used computer vision and explicit validation so a sports talent platform can assess a recorded drill consistently — starting with football, with the same pattern ready for the next drill.

Have a prototype that looks right and scores wrong?

Tell us what the athlete records, and what a fair score has to mean. We'll tell you what the first drill should prove.

All case studies