Case Study - A 10,000+ Polish recipe database, built with an AI pipeline
A database of over 10,000 publication-ready Polish recipes - curated from public-domain data, translated with full editorial adaptation, categorized, illustrated with AI imagery and guarded by automated quality gates.
- Client
- LosSmaku
- Country
- 🇵🇱Poland
- Year
- Service
- AI content pipeline
Overview
LosSmaku is a culinary service built around a randomizer that answers the eternal question "what's for dinner?". To launch, the client needed a large, credible base of Polish-language recipes - and writing thousands of recipes by hand was beyond both budget and timeline.
The constraints were hard: public-domain data only, content that reads editorially rather than machine-translated, and delivery in a CSV structure matching the client developer's system.
How it works
- Selection - from a public-domain (CC0) dataset, the pipeline picks only worthwhile recipes: a popularity ranking weighted by ratings, only entries with a genuine author's description, automatic rejection of incomplete ones.
- Translation with full editorial adaptation - a glossary of localized ingredients, metric units with kitchen-friendly rounding, US brand removal, titles generated fresh from the recipe's content instead of translating the original.
- Categorization, servings and AI imagery in the brand's style and colors - all mapped to recipes directly in the delivery files.
- Automated quality gates - step counts checked against the source, detection of leftover English, scans for brands and imperial units; anything that failed went back for fixing before the client ever saw it.
- Live preview - a simple web app where the client approved quality before receiving the files: first a pilot and a feedback round, then the full database.
The result: a complete, consistent Polish recipe database - with descriptions, categories, servings and images - ready to deploy in the client's service.