Lecture Video Pipeline
A pipeline that turns a presenter recording and a slides recording into one branded 1080p lecture video - intro, outro, partner branding and Whisper subtitles - and uploads it to the video platform. Built during my rotation in Switch's video team.
- Type
- Work · Switch · Pipelines / Speech AI
- Year
- 2026
- Role
- Developer in the VIDEO team - from the first CLI to the production service.
Cut render time from 34 to 16 minutes with parallel ffmpeg rendering, fixed an out-of-memory crash (7+ GB down to a constant 1.2 GB), and shipped it to production for a real university course.
- 34 → 16 min render time, A/B tested
- 7+ GB → constant 1.2 GB memory
- In production for a real university course
The problem
Lecture recordings came as two separate videos - the person and the slides - and putting them together with branding and subtitles was slow manual work for every single video.
One job that takes the raw recordings and returns a finished, subtitled, branded video on the platform, fast and without babysitting.
What I did
- Built the compositor as a CLI first, then grew it into a FastAPI service with a queue worker and automatic upload to the video platform.
- Moved rendering from MoviePy to direct ffmpeg calls, split long videos into segments that render in parallel, and proved the gain with an A/B test.
- Found why long renders crashed: 8-9 overlay inputs pushed memory past 7 GB. Merging them into one input kept it at a constant 1.2 GB.
- Deployed it on a Linux VM with Ansible and systemd, with health checks and onboarding docs for the team.
Outcome
In production: renders 2.1x faster (34 to 16 minutes), a full job dropped from 82 to 14 seconds, and it delivered 18 videos with subtitles in four languages for a university course.
Stack
- Python
- FFmpeg
- Whisper
- FastAPI
- Queue worker
- Ansible
- systemd