An AI highlights pipeline for social video
An automated pipeline that watches a live sports stream, uses AI to find the key moments, cuts them into vertical clips and publishes them to social — with no human editing in the loop.
We used AI for exactly the part that's hard to automate — spotting the newsworthy moment — and deterministic tooling for the rest. The result is a pipeline that's fast, cheap to run and genuinely hands-off.
This platform turns a live sports stream into social-ready highlight clips automatically. It records the feed, transcribes the commentary, and uses an AI model to find the moments that matter — goals, penalties, cards — then cuts each one into a vertical clip and publishes it, all without a human editor.
We built it on .NET as an event-driven, queue-based pipeline: a live stream is segmented, its audio transcribed, the transcript analysed by a language model with a strict function-calling schema, and each detected moment reframed from wide to vertical with GPU-accelerated encoding before being pushed to social.
It watches the match and posts the highlights before we'd have finished exporting the first clip by hand.
A multi-stage approach.
Capture
We resolve a live stream and record it into rolling short segments, on a timer that can wait for a scheduled kick-off before it starts.
Transcribe
Each segment's audio is transcribed to a timed transcript — with both a hosted API and a local GPU model trialled for speed and cost.
Detect
The transcript goes to a language model with a strict schema that returns the key moments as structured data — goals, cards, penalties — each with precise start and end times.
Reframe & publish
Each moment is cut, reframed from 16:9 to a 9:16 blurred-background clip on the GPU, then uploaded and published to social with an auto-generated title and hashtags.
What changed.
AI watches the game
The system finds highlights from the commentary, not the pixels — a language model reads the transcript and returns the key moments as structured, timestamped data.
Vertical, automatically
Every clip is reframed from wide broadcast footage to a 9:16 social format with a blurred backdrop, encoded on the GPU for instant mobile playback.
No editor in the loop
From live feed to published Short, the whole pipeline runs itself — turning minutes-old action into posted clips.
Built to reprocess
An event-driven, queue-based design means any stage can be re-run on its own, so a failed step never means recapturing the stream.
Thinking about AI or media automation?
If you've got a manual, repetitive media or content process, we'll help you work out where AI genuinely fits — and build the pipeline around it.