Fifty-seven clicks, one drag
For sports photographer Uğur Şenpire’s team, I built a cloud pipeline that turns a selected match photo into finished, branded 4:5 and 9:16 posts, ready in seconds after it lands in a Drive folder.
- Year
- 2026
- Role
- Sole builder and maintainer
- Scope
- Rule extraction from the editor’s own files, color, cropping, watermark and sponsor placement, Drive integration, cloud deployment, blind tests with the photographer
- Stack
- Python, pose detection (RTMO) and object detection (YOLOX), 3D color LUTs, Google Cloud Run, Google Drive API, Photoshop Camera Raw for reference targets


Same frame, same 4:5 window. Left, the camera’s file; right, the finished post with the team’s color, watermark and badge.
Uğur Şenpire photographs football matches, and his team posts the best frames while the match is still fresh. Every post meant the same routine in Photoshop: grade the color, crop for the feed and again for stories, place the studio’s watermark, add the sponsor’s badge. Recreated step by step, one photo took fifty-seven clicks. The goal I set was a finished frame within two or three minutes of the photo being chosen.









UGUR0732.JPG

Real Photoshop screens and the pipeline’s real files for the same frame. Typing is not counted, so 57 is the floor.
I didn’t guess the rules; I measured them. The editor’s Photoshop files kept every crop box and color setting, so I read them back and rebuilt the logic from there. Crops follow the main player’s body. The watermark sits on plain fabric of the shirt, never on a face, skin or printed text. Pose detection finds the players and ignores fans in the stands. Folder names choose the recipe: full system, color only, or your own crop.
4:5
The crop windows and the watermark spot are the pipeline’s own, read from its log. The boxes for face and print are mine.
Two decisions changed the project. The first version copied the old editor’s color recipe faithfully, and the team disliked it: the grass turned cyan. From then on the reference was what the team approved, and every color change went to a blind test instead of a score. In the latest one, Uğur picked the team’s own look in all twelve night frames. The second came from Uğur zooming into a face. Every output had been scaled down to 1080 pixels, and the skin went grainy. Now each frame keeps its native resolution.
Since October 2026 the pipeline has run on Google Cloud in Uğur’s own Drive. A live series of five photos finished in fifty seconds. Crops overlap the editor’s own by 0.88 on average. When a frame can’t be handled with confidence, the pipeline names the reason in the file name and leaves it for a person. Each release passes 74 end-to-end checks, and a second AI model reviews the code before it goes live.







Working on something like this? Tell me what’s missing.
NextEuroLeague Women Store