Developer shares hard-won lessons from building a 12-language AI product search tool
A developer building OneFindMe, an AI-powered product search frontend, deployed the entire system on Cloudflare Workers using Claude Haiku for query translation and image recognition. One key challenge was that LLM-generated translations produced natural prose rather than the keyword-style strings that marketplace search engines expect, requiring a prompt redesign. The system also faced issues with the AI confidently returning non-existent category IDs, which was resolved by validating model output against a strict allow-list and treating category suggestions as optional enrichment rather than core logic. To address slow cold-path searches — which involve multiple LLM and API calls — the team implemented instant bestseller previews and pre-caching of new search terms to mask latency. The project now supports 12 languages, with each bug surfacing only after real multilingual users began searching.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in