Multilingual Video Localization Pipeline
How an automated subtitle pipeline reduced video localization from 3 days to 15 minutes across 12 languages.
Enterprise Software, 3 weeks
The client
A global technology company localizing product videos, webinars, and marketing content into 12 languages for European and Asian markets.
The challenge
The localization team was manually translating subtitles for product videos. Each video took 2-3 days per language. With 12 target languages, a single video could take weeks to fully localize. Quality was inconsistent: translators working independently produced different terminology for the same product features. Glossary compliance was impossible to enforce at scale. The team had a growing backlog of untranslated content, and every product launch created a bottleneck.
The approach
Built a four-stage automated pipeline. Stage 1: Claude AI pre-merges fragmented subtitle blocks into complete sentences, fixing the broken fragments that SRT format creates. Stage 2: DeepL translates with glossary enforcement and per-language formality rules (formal for German, French, Japanese; informal for Spanish, Italian). Stage 3: Claude condenses any translations that exceed the 2-line subtitle limit for screen readability. Stage 4: Timestamp splitting as a final fallback for edge cases. The pipeline processes all 12 languages simultaneously. A companion Google Apps Script version automates batch processing directly from Google Drive.
Built with Next.js, Claude AI (Haiku), DeepL API, Google Apps Script, and Kubernetes.
How the engagement ran
Observe
Mapped the manual translation workflow. Identified the root cause of quality issues: fragmented SRT blocks breaking sentence context.
Decompose
Separated the pipeline into four independent stages, each handling one specific failure mode.
Build
Web app for on-demand translation plus Google Apps Script for batch automation. Both deployed and used by the localization team.
Codify
Documented per-language rules (formality, glossary, formatting) so the system encodes institutional knowledge that previously lived in translators' heads.
The results
Localization time dropped from 2-3 days per language to 15 minutes for all 12 languages simultaneously. Glossary compliance reached 100% through API-enforced terminology. Per-language formality rules are applied automatically. The backlog cleared within two weeks of deployment. Japanese subtitles follow specific formatting rules (half-width spaces instead of commas, full-width spaces instead of periods) that were previously handled inconsistently by human translators.
Localizing content at scale?
Tell us about your operation. We will tell you what a system like this would take to build.