Automating LLM translation cleanup for product data
Cleaning up product translations inside Windly
Context
Windly surfaces overseas products to domestic buyers, so translation quality is product-page quality.
Option names, titles, and tags were each translated their own way, so the same product went by different words on different screens.
How I solved it
- Switched the approach from rule-based to LLM-based cleanup, designing and building the pipeline on Langchain
- Proved it internally with a Streamlit MVP first, then built the serving backend
Result
- The cleanup pipeline shipped to production inside Windly
- Applied across roughly one million products
- option names, titles, and tags