Challenge
Short videos for several brands, in two languages and on several platforms, mean a lot of repetitive work: script, graphics, voiceover recording, subtitles, editing and publishing. The previous generator was a single application on a single server and served one brand.
We wanted a content factory that scales to more brands and languages, but publishes nothing without human consent and grows only organically.
Approach
The domain model is brand, language version and channel. An episode does not depend on language: the English version is an adaptation of the script, and illustrations are reused, which significantly lowers the cost of the second version. Formats are data-driven templates, so a new format needs no changes to the pipeline.
We compute the render on the existing cluster instead of a separate server. Measurements showed that at the current scale a dedicated server would sit idle over 90% of the time. The thresholds at which buying one makes sense are written down in advance.
Implementation
The pipeline: a language model writes the script, an image model creates the illustrations, a speech synthesiser records the voiceover with voice settings per language version, subtitles are built from per-word timings, and the render goes through a job queue. A quality control stage catches faulty files, and episodes with no active job after 90 minutes are marked as failed.
The team approves episodes in a PWA on the phone, with notifications about finished material and production errors. Publishing goes through a swappable adapter with a circuit breaker, and a dashboard shows costs and render load.
Results
After optimisation, an episode in two language versions is produced in 181 s instead of 219 s, the render uses about 45% less CPU, and the memory reserved for the worker dropped from 3 to 1 GB. With the shorter render, image quality went up (SSIM from 0.971 to 0.985).
First production run: 3 videos in 4 minutes 50 seconds, with no errors, for 0.70 USD, or about 0.23 USD per video. All three were published on both target platforms.
