DoubleTake
Recruiting-video platform for athletes and families, with a serious pre-ChatGPT computer-vision POC running behind it. The AI never reached the live product. The April 2020 proof was real: 74 annotated videos, 12.6M player crops, 92.8% Re-ID validation triplet accuracy.
- Apr 2020
- internal AI POC
- 12.6M
- player crops generated
- 92.8%
- Re-ID validation triplet accuracy
- 1,031
- registered users
- Six figures
- outside capital raised
Game film in. Reel out.
A MERN platform served the recruiting-video product. An Active Player Tracking POC ran underneath it. The domain is still live.
Four steps: raw game film in, a shareable reel out. The premise was that a tracker could find the athlete for the editor.
A vision POC meant to hold onto one athlete across occlusion, multiple cameras, and bad lighting, then isolate only the moments of action.
- INPUTRaw sideline or broadcast footage uploaded per game by the parent, tagged once with the athlete's jersey number.
- ANCHORRe-identification anchored to the jersey number, so one athlete stays tagged across every play and angle.
- TRACKA Siamese triplet Re-ID model holds the athlete through occlusion, crowding, and camera switches, where ordinary tracking loses them.
- SEGMENTPose keypoints feed a compact action model that flags moments of action rather than dead time.
- PROOF92.8% Re-ID validation triplet accuracy on 12.6M player crops from 74 annotated videos.
- OUTPUTA runnable demo path from athlete tag to highlight candidates.
What the tracker actually saw.
Raw game film in; detection, tracking, pose, and action scoring out. These are real frames and a 31-second clip from the April 2020 computer-vision build, running well before the generative-AI wave.
What it runs on.
The shippable evidence.
The recruiting-video workflow was the public surface. The April 2020 vision POC's own output is shown below; the full report, archived code, and financials stay private.
doubletake.video. Public DoubleTake product shell.
publicCV/ML system: player detection, CSRT tracking, Siamese Re-ID, pose estimation, and action detection, used to find highlight candidates in game film.
demo output shown on-page · system stays privateInternal April 2020 delivery, plus a running-action detector tuned to an 80% recall threshold.
private · technical notes and archived repoSix figures of outside capital raised. Organic growth to 1,031 users across athletes, parents, and coaches (no paid acquisition). Roughly $30K of revenue during the COVID year.
public launch evidence starts 2018 · paid product traction through 2022