Clips Kitty Game Day

Clips Kitty Game Day turns live sporting moments into social media posts in minutes, not hours, with Meta's models doing the work and a person approving every post. I'm a Toronto Raptors fan, and I was inspired by the Instagram page House of Highlights. Imagine you're the social media manager for an NBA team. You're at the game, not at an editing desk, and the clips need to go out while people are still talking about it. Usually the footage gets reviewed, an editor cuts it at a computer, and the posts go out hours later, after the buzz. Game Day gets them out while the game is still trending: more reach and ad value for the team, the manager stays at the game, and more teams can post more Reels on Instagram and Facebook. How it works: 1. The first quarter lands in a shared folder, or is uploaded on the dashboard. The server picks it up on its own. 2. Muse Voice Transcribe writes down what the announcers say, with times. 3. Muse Spark picks the plays worth posting, held to a strict JSON schema, and writes each clip's title and caption. 4. SAM 3.1 finds the ball through each play, so the 9:16 crop follows the action. ffmpeg cuts the clips. 5. Your phone buzzes with a line Muse Spark wrote. Ask your Muse agent what's ready, and it hands you the approval link through a custom connector I built. 6. On the approval page you watch the clips, pick Instagram and Facebook, confirm, and enter your PIN. Only then do they go out as Reels through Meta's Graph API. Only Meta's models run AI in the app: Muse Voice Transcribe, Muse Spark and SAM 3.1, through the Meta Model API. Everything else is a regular service: an AWS Lightsail server, a Cloudflare Worker relay that keeps the Instagram and Facebook tokens off the server, Google Drive for the shared folder, ntfy for phone alerts, and ffmpeg. What has run: a 45-minute game made 8 clips in about 14 minutes on a $24-a-month server, for about $0.80 of Meta usage. On 4 Nov, a 4-minute piece of a game was dropped into the shared folder. 4 clips were ready about six minutes after the upload, and one was approved on my phone with the PIN and posted to Instagram and Facebook (@clipskitty2). Something surprised me: in an offline test, SAM 3.1 found the ball about twice as often as E-BARD, a YOLOv8 model trained on basketball (github.com/GabrieleGiudic/E-BARD, used unmodified under CC BY 4.0, offline only). Next: fine-tune Muse Glimmer on basketball footage, add WhatsApp alerts (Meta locked my new business account before the first message went out), and post to Threads. Built during the event: everything in the repository. Clips Kitty, my earlier desktop clipping app, isn't part of this entry. PROVENANCE.md lists where every file came from. Demo footage: DVIDS / Justin Fairley, 2D Audiovisual Squadron (public domain). The appearance of U.S. Department of War (DoW) visual information does not imply or constitute DoW endorsement. Demo video: https://youtu.be/1_QXVlY6cqE Code: https://github.com/ColinGPT9/clips-kitty-gameday Try it, no account needed: https://18-190-220-192.sslip.io (user: judge, password: kfFHEPCBLT3nO0dMO87Fq4qy)