The client is a digital publishing network running regional news sites across several Asia Pacific markets, previously relying on a mix of manual translators and a legacy machine-translation vendor whose output frequently missed cultural and editorial nuance. However, they faced significant hurdles:
- Translation Backlog: Only a fraction of daily articles could be translated in time to stay relevant for breaking news cycles.
- Quality & Nuance Loss: The legacy machine translation vendor produced overly literal translations that missed idiom, tone, and cultural context.
- Cost of Human Translation: Fully human translation across six languages at their daily article volume was financially unsustainable.
- Fact Drift Risk: Any AI translation system needed hard guardrails against silently altering names, numbers, or direct quotes.
Varcio built a localization pipeline on Gemini with fact-preservation as a non-negotiable design constraint from day one:
- Gemini 3.5 for Translation: Fine-tuned prompting with regional style guides tailored to each language and market's editorial conventions.
- Fact-Preservation Guardrails: A validation layer diffed named entities, numbers, and quoted text between the source and translated article, automatically flagging any mismatch for mandatory human review before publication.
- Vertex AI Pipelines: Orchestrated the full ingest, translate, validate, review, and publish workflow automatically for every article, with no manual handoffs.
- Regional Editor Review Queue: Local-market editors did a fast pass-review of AI translations rather than translating from scratch, cutting per-article review time dramatically.
- Glossary & Style Memory: A living glossary of brand terms, public figures' names, and regional style preferences that improved automatically with every editorial correction.
Localization Time Per Article (Minutes)
Validating translation quality and editorial workflow on two languages before wider investment.
Building the fact-preservation diffing layer that flags entity, number, and quote mismatches.
Onboarding regional editorial teams onto the new review-and-approve workflow, market by market.
Scaling the pipeline to all six languages across every regional site in the network.
Translation costs dropped by 60% even as article coverage went from partial to near-complete across all six languages. Breaking-news localization time fell from hours to minutes, letting regional sites publish translated coverage while stories were still developing instead of days later.
Factual accuracy incidents — the network's biggest fear going into the project — stayed near zero thanks to the guardrail layer, which caught entity and number mismatches before they ever reached a live article.
"We used to choose which stories were worth translating. Now we translate everything, and our editors spend their time on judgment calls instead of retyping sentences."