Design an LLM-powered localisation pipeline that translates product content into 20 languages.

30-second answerSay your answer out loud first, then reveal.

Extract translatable strings and content with context (screenshots, descriptions, length limits, placeholders), and apply glossaries and translation memory first (reuse approved translations). Translate with an LLM using per-language style guides and terminology, then run automated QA (placeholder and format integrity, length limits, terminology compliance, back-translation or LLM quality scoring). Route low-scoring or high-visibility content to human linguists, feed edits back into translation memory and few-shot examples, and integrate with CI so new strings ship localised.

Pipeline

Extraction: strings from code and CMS with keys, context notes, character limits, placeholders ({name}, HTML tags, plurals).
Translation memory (TM) lookup: exact or fuzzy matches reuse approved human translations (consistency + cost).
LLM translation: prompt includes glossary terms (brand names not translated), style guide per locale (formal vs informal "you", e.g. "aap" vs "tum" in Hindi), surrounding strings for context. Batch similar strings together.
Automated QA:
• Deterministic: placeholders preserved, tags balanced, length within UI limits, numbers, dates and currency formats.
• Terminology check against the glossary.
• Quality estimation: LLM judge or COMET-style metrics; back-translation comparison for meaning drift.
Human-in-the-loop: linguists review content flagged by QA plus high-visibility content (legal, marketing, onboarding); sampled audits for the rest.
Feedback: edits update TM and glossaries; recurring corrections become few-shot examples per language.
Delivery: CI integration (a new string triggers translation and QA; a pull request with translations); visual in-context review with screenshots.

Language-specific concerns: right-to-left scripts, Indic scripts' rendering and length, pluralisation rules, culturally appropriate examples, and quality differences across languages (evaluate per locale).

Metrics: cost and turnaround per word, human edit rate per language, QA failure rates, user-reported translation issues.

PreviousDesign a self-service fine-tuning platform for internal teams (upload data → train → evaluate → deploy).

Slow is fine. Stopping is the only problem.

Design an LLM-powered localisation pipeline that translates product content into 20 languages.

Pipeline

Related