How I built Abra, an AI creator crew on Meta's Muse models

When a Facebook video takes off, the work around it can cost a creator a whole team. In my biggest years it cost up to $150,000 in help. Abra is my attempt to automate that work.
What it does
• Muse Voice transcribes each video.
• Muse Spark writes three titles, a caption and hashtags.
• Muse Wonder picks the hook and explains why.
• The chat answers from live page numbers, and Creator Spotlight suggests ideas for a genre.
• SAM 3.1 cuts sticker cut-outs from a selfie.
• Muse Spark screens comments, and Claude drafts in-voice replies.
How it's built
• A Python standard-library HTTP server with a single-page dashboard
• SQLite for state, ffmpeg and Pillow for frames and captions
• Meta's Responses API for the Muse models
• The SAM 3.1 mask decoder ported to pure Python
What went wrong
• Every Meta call returned 402 until billing was set up, so the app ran on Claude in the meantime.
• The Baseten open-weight path still returns 402, so it doesn't run.
• A billing error turns Muse off until the app restarts. That's a weakness I'd fix before production.
Demo

Code (MIT): https://github.com/amyboissoneau/AbraMeta