AI translation + AI agents = i18n made easy
summary
Shows that chaining LLM steps can improve translations but is slow and costly: 34 seconds for a single line and roughly 3,680 tokens spent when prompting Gemini 2.5 Pro. A 4‑agent pipeline (translate, fluency, adequacy, style, then editor) scores worse than Google Translate and DeepL in one paper; DeepSeek R1 does slightly better. Demonstrates an agent that scans React files, creates keys and T() calls, writes JSON locales and uses tools like DeepL API and Google Sheets MCP, but requires human review at every stage.
AI-written summary · about
chapters (4)
- 0:00AI translation and agents
- 3:55Improving AI translations
- 8:11Internationalizing website
- 16:45Automating translation workflows
related talks
50mLet's build an AI agent
18mBuilding an AI service without training – AI translations with LLMs
39mBuild the new generation of AI Agents
10mBuilding Conversational AI Agents at Scale
01h 48mBuilding Conversational Agents
01h 45mAI Engineering with the Google Gemini 2.5 Model Family
43mBreaking the language barrier: how do we quickly add multilanguage support in our AI application?
36mProduction-Ready AI Agents: From LLMs to Small Language Models