AI Video Pipeline Clipping
An internal AI video platform that turns long videos into short, captioned clips or highlight reels.
The Problem
Turning a 90-minute match broadcast, stream or talk into publishable clips meant hours of manual scrubbing, cutting, reframing and captioning. Generic clippers also treat every video the same, so football highlights, podcasts and music need different detection and editing logic.
What I Built
I designed a queue-based system where a NestJS API hands work to BullMQ workers. They download the source, transcribe it with Deepgram, and use an LLM to find and rank the strongest moments. FFmpeg then cuts, reframes and captions each clip, and Supabase Realtime streams progress to the React dashboard. Seven pipelines share this base, each with its own detection and editing strategy. When the football pipeline invented facts such as wrong player names, I moved everything checkable out of the model and into deterministic rules. Access is invite-only with admin and editor roles, row-level security and an audit log. An AI coding agent wrote the application code from my specs, and I owned the architecture, database, infrastructure and verification.
The Result
100+ clips produced from full 90-minute match broadcasts 18.5K views and 46.7 hours of watch time on published clips Used by a small internal team for real publishing jobs, with admin and editor roles Seven pipelines live: generic, football, motion graphics, player compilation, music, movie and stream Clip and reel output in four selectable aspect ratios, with live progress in the dashboard