Global recipes, local voice
Multilingual culinary translation service built to scale editorial quality across 11 languages.
Use case. Domain-tuned LLM for industrial translation of culinary content for a global small-appliance group.
Scope. 11 languages, several thousand recipes, 150+ countries.
Status. In production, used daily by marketing teams.
Context
Our client is a global player in culinary small appliances: its products are used in kitchens across more than 150 countries. Each product line (food processors, air fryers, bread makers, yogurt makers, etc.) comes with a large editorial footprint: recipes, cooking instructions, and preparation guidance.
At group scale, this means several thousand recipes to maintain, enrich, and translate into 11 languages, in sync with product launches and catalog updates.
Three facts triggered the mission:
- editorial volume keeps growing as product ranges expand;
- generic translation tools are not reliable enough for culinary language;
- editorial consistency degrades across batches and countries.
The goal was not to plug in machine translation, but to build an industrial service that translates accurately, preserves brand voice, and matches enterprise publishing speed.
Problem to solve
A generic translation engine fails in this context:
- it translates literally where culinary verbs require precise culture-specific equivalents;
- it ignores units, producing recipes that are unusable for target markets;
- it cannot handle ingredient substitutions for local availability;
- it does not carry brand voice (tone, register, style);
- it does not accumulate learning: every batch starts from scratch and terminology drifts.
Without a dedicated solution, teams face heavy human review cycles, slower go-to-market, cross-market inconsistencies, and rising editorial debt.
Our solution
We designed and deployed an end-to-end multilingual culinary translation service built on four components.
1. Structured multilingual culinary corpus
Technical terms, cultural equivalences, domain glossaries, unit conventions, and market-specific ingredient substitutions are centralized, structured, and versioned. This corpus becomes the service's domain memory.
2. Cuisine-specialized LLM
Instead of a generic engine, we fine-tune an LLM for culinary context. The model learns domain terminology, usage constraints, unit handling, and when a local substitution is better than literal translation.
3. Industrial script library
We provide a script library that industrializes batch processing, runs automated quality checks, handles unit conversions, and generates auditable outputs integrated with the client's existing tooling.
4. API and business-facing web application
- The API is callable from client tools (CMS, DAM, product platforms): translation becomes editorial infrastructure.
- The web app allows marketing teams to submit, review, and validate translations in a clear workflow, without depending on engineering teams.
What changes for the client
- Editorial quality. Translation output is directly usable, with far less manual correction.
- Time-to-market. Product launches are no longer blocked by translation backlogs.
- Global consistency. Techniques, ingredients, and units are translated consistently across markets.
- Scalability. The service handles large volumes without quality decay.
- Editorial sovereignty. Corpus, glossaries, and conventions remain client owned.
Principles behind the result
- Domain first, technology second. The solution is built from real culinary translation needs.
- Corpus as a strategic asset. Value comes from the model + domain corpus pair.
- Human-in-the-loop by design. Marketing teams validate and improve output continuously.
- Industrialization from day one. API, scripts, and app cover technical, bulk, and editorial workflows.
Capabilities mobilized
This project combines:
- domain knowledge engineering (terminology, glossaries, equivalences);
- domain LLM adaptation for culinary translation;
- service architecture (API, library, web application);
- editorial UX for non-technical teams;
- MLOps (batch pipelines, quality controls, traceability).
A reusable model
The same approach applies to any industry where high-volume translation must respect domain precision: technical manuals, standardized product sheets, regulatory content, or multilingual editorial assets.
The method stays consistent: structured domain corpus, specialized LLM, industrial toolchain, and workflows designed for operational teams.