Lengua Fina

A transcreation agent for translators and small producers who sell abroad.

Translation keeps the words and loses the references. "Smooth as a Brussels praline" works in Brussels and means little in Tokyo. Lengua Fina rewrites product copy for each target market using cultural references that audience actually has affinity for, according to Qloo's taste graph, and flags the source copy's references that won't land.

In Spanish, lengua means both language and tongue: the thing you speak with and the thing you taste with.

How it works

An agent (Claude Opus 5.5, via the Anthropic SDK tool runner) gets the copy, the product and up to three markets (city, age ranges, language). It has five tools:

Tool	Qloo endpoint	Purpose
find_entity	/search	Resolve references in the source copy to Qloo entities
find_tags	/v2/tags	Find product tags (e.g. chocolate) to narrow an audience
market_taste	/v2/insights	What an audience (city + ages + interests) has affinity for, per entity type
check_reference	/v2/insights	Whether a given reference ranks in the top 100 for an audience
submit_localization	–	Final copy per market, with references cited by Qloo entity id

Grounding check: submit_localization rejects any reference whose entity id was not returned by a Qloo tool during that run, and sends the agent back to fix it. Every reference shown in the UI links back to real Qloo data.

Run locally

cp .env.example .env.local   # add QLOO_API_KEY and ANTHROPIC_API_KEY
npm install
npm run validate             # checks that Qloo returns different taste profiles per market
npm test                     # grounding check unit tests
npm run dev

npm run validate -- Tokyo Brussels "Mexico City" compares any markets you like. Raw Qloo responses go to scripts/out/, which is gitignored: Qloo data must not be committed.

Public demo limits

The live demo allows 6 runs per IP per hour and 300 runs a day (RATE_LIMIT_PER_HOUR, RATE_LIMIT_PER_DAY).

Stack

Next.js 15, TypeScript, Tailwind, @anthropic-ai/sdk, Zod. Deployed on Vercel.

License

MIT