Skip to Main Content

kim & yang

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

kim & yang — Demo video

AdLoom helps small businesses turn a phone video into advertising content they can review and control. The application accepts a 3–45-second video and a short typed or spoken brief, then prepares six video variants: three aspect ratios (9:16, 4:5 and 1:1), each with two hooks. Owners can preview the results, inspect product-following crop paths, edit copy and scene selection, review factual claims and policy warnings, and approve a specific revision. The architecture integrates Muse Spark for structured video understanding and copy, SAM 3.1 for product tracking, Muse Image for still thumbnails and end cards, and Muse Voice for transcription. Local ffmpeg rendering assembles the videos without AI-generated motion. Publishing is designed around explicit human approval. A SHA-256 fingerprint covers the reviewed content, and edits invalidate that approval. The Marketing API integration creates campaigns, ad sets and ads only in PAUSED status. A persistent request journal blocks automatic retries when a publishing outcome is unknown. AdLoom includes a responsive FastAPI web app and CLI, with 520 automated tests passing in the latest full verification. The demonstrated media workflow uses synthetic footage and fixture providers; the film clearly labels its Figma design previews. Live Meta authentication, model listing and a minimal Muse Spark text completion have succeeded. Full live media processing and real ad creation remain unverified.

Team