DoubleTake Rebuild
The DoubleTake rebuild is the AI player-tracking product I originally set out to make. It follows one athlete through a full soccer match, finds the moments they are involved in, and turns those clips into a reel the family can review and edit.
- 1 athlete
- tracked through the match
- 6 stages
- from upload to finished reel
- No face scans
- body, kit, and jersey drive identity
- 5 clips
- golden-corpus release gate
An AI player-tracking system built to follow one athlete through a full soccer match and turn their moments into an editable reel.
DoubleTake RebuildThe AI follows one athlete through the game.
The current product moves from a resumable upload to one athlete-confirmation tap, a six-stage tracking run, and a reel the family can trim, reorder, and export. These frames come from the canonical July 27 build spec and map to implemented app surfaces.
Upload
Upload a full match from a phone, tripod, Veo, Trace, Hudl, or another camera. Two halves can travel as separate files, and an interrupted upload can resume.
resumable · multiple filesConfirm
DoubleTake finds the players and presents a crop gallery. The family confirms the athlete using jersey and kit cues, without facial recognition.
photo grid · body and kitTrack
The tracking pipeline follows that athlete through crowded play, camera movement, and temporary occlusion. If identity gets uncertain, the system asks for another confirmation.
player identity · full matchBuild the reel
Detected moments become a first cut. The family can keep, trim, drop, reorder, or add a missed clip before the reel is rendered again.
EDL · preview · exportThe working parts behind tracking, confirmation, and reel review.
The system tracks a player, not a face.
Athlete identity comes from what is visible on the field: body shape, kit, jersey number, and motion. Facial recognition is off the table.
body · kit · jerseyUncertainty becomes a question.
When the tracker is no longer sure it has the right athlete, it surfaces a small set of crops and asks the family to confirm before continuing.
up to three confirm tapsEvery detected moment stays editable.
The first cut is a starting point. Each clip can be adjusted or removed, and a missed moment can be added from the game footage.
keep · trim · drop · addDifficult footage is part of the test set.
A small golden corpus covers the conditions that usually break a tracker. Pipeline changes have to clear those clips before they reach a real match.
occlusion · crowding · camera motionThe tracker has to be useful without pretending it is certain.
- Launch copy describes soccer Game Edits only.
- An adult account owner must confirm they are authorized to process the footage.
- Athlete identity uses body and kit evidence. Facial recognition is prohibited.
- The family confirms the athlete before the full tracking run continues.
- A queued or uncertain pipeline job is never reported as a finished reel.
- Browser preview and exported video read the same versioned EDL contract.
The web product, the tracking pipeline, and the hard clips used to test it.
The Next.js product surface for upload, athlete confirmation, tracking status, reel review, editing, and billing.
private beta codebasePython stages for ingest, player detection, athlete identity, tracking, moment detection, and rendering.
Python · Modal · FFmpegA small consented corpus covering crowded play, occlusion, camera changes, poor lighting, and handheld footage. Pipeline changes are measured against it.
quality gate