A slow, editor-by-editor Premiere Pro process turned into an automated pipeline that cleans, reframes, and AI-edits recorded lesson videos in minutes — built to run hundreds of videos at volume.
Black Camel Productions produces a high volume of recorded lesson videos for an educational content client — screen-and-webcam recordings of a teacher walking through slides. Every video went through the same repetitive manual edit in Premiere Pro:
On top of that, an editor had to listen through each video to remove silences, filler words, coughs, and repeated or stumbled lines — the most time-consuming part of all.
The problem wasn't quality — it was throughput. Each video cost an editor roughly an hour of hands-on time, the process didn't scale, and quality drifted between editors. At volume, the manual content-cleanup pass became the bottleneck.
We built an automated video-production pipeline that reproduces the manual editing standard programmatically — plus an AI-assisted content-cleanup layer the manual process couldn't do efficiently.
It runs as a managed service: Black Camel sends the raw recordings, the pipeline processes them, and a finished, brand-consistent video comes back — ready for their team to drop into final assembly.
We reverse-engineered the client's exact editing standard from their original Premiere presets and reference videos, so output matches their spec to the pixel — webcam framing, spacing, and scale measured and locked (for example, a 42px webcam margin matched against a reference video, not guessed).
Toolbar and UI removal, page-break masking, screen re-framing, and presenter-bubble enlargement happen automatically — no manual masking or keyframing.
The system inspects each slide and adjusts automatically so on-screen images and tables are never cut off and never collide with the presenter overlay — solving real client feedback without per-video manual work.
Brand geometry was decoded from the client's own design assets and measured against a reference video, not estimated — so output is consistent and on-spec every time, eliminating editor-to-editor drift.
For finding repeated lines, the pipeline pairs Whisper transcription (with word-level timestamps) with a text-matching layer tuned for the task — locating duplicated speech precisely and proposing a clean cut, with a human approving the final call.
We tuned the render engine for Apple Silicon with smart frame-reuse, roughly halving render time — from about 6 minutes down to ~3 minutes for a typical 15-minute lesson.
Batch processing with automatic retries on network hiccups, persisted results, and visible warnings so nothing fails silently across a large batch.
Black Camel moved from a labour-bound, one-editor-per-video process to a scalable production line: send the recordings in bulk, let the pipeline do the heavy, repetitive work, and apply human judgement only where it adds value.
The result is lower cost per video, faster turnaround, and consistent brand quality at volume — turning video production from a bottleneck into a throughput advantage.
Delivered for Black Camel Productions; the educational end-client is kept anonymous by agreement.
| Language | Python |
| Video processing | FFmpeg · OpenCV |
| API / Orchestration | FastAPI |
| Transcription | Whisper (word-level timestamps) |
| Speech cleanup | Cleanvoice |
| LLM / Text analysis | Gemini |
| Optimisation | Apple Silicon render tuning |