Kaikaku.AI Introduces Epicure: AI Models Differentiating Culinary Relationships from Chemical Pairings

London-based startup Kaikaku.AI has unveiled "Epicure," a set of AI models that differentiate between ingredients frequently paired in recipes and those sharing similar chemical flavor profiles. This distinction, previously often blurred in AI models, could impact how culinary professionals and food tech companies approach recipe development, ingredient substitution, and menu innovation. The models, developed by Jakub Radzikowski and Josef Chen, were trained on a vast dataset of 4.14 million recipes across seven languages and the FlavorDB chemical database.

The Epicure models – "Cooc," "Chem," and "Core" – each offer a distinct approach. Cooc learns from co-occurring ingredients in recipes, while Chem focuses solely on shared flavor molecules. Core blends both methodologies. This separation yields varied recommendations; for "chicken," Cooc suggests garlic and onion, common recipe partners, whereas Chem proposes beef or pork, indicating similar flavor profiles. This nuanced understanding can be particularly valuable in a country like India, with its diverse regional cuisines and complex flavor combinations.

Key facts:

Feature	Description
Developer	Kaikaku.AI (London-based startup)
Model Name	Epicure (includes Cooc, Chem, Core)
Training Data	14 million recipes (7 languages) + FlavorDB chemical database
Key Innovation	Differentiates recipe-based pairings from chemical flavor relationships

Unpacking the Models

The "Cooc" model, trained on recipe co-occurrence, provides suggestions based on traditional culinary pairings. For example, a query for "basil" might yield parsley, olive oil, and parmesan, reflecting typical pasta dish ingredients. In contrast, the "Chem" model, leveraging the FlavorDB chemistry database, identifies ingredients with shared flavor molecules. For "basil," Chem suggests oregano, tarragon, and rosemary – herbs with similar underlying chemical components.

Interestingly, the chemistry-driven "Chem" model demonstrated an unexpected capability: it classified flavors (sweet, sour, bitter) and nutritional values (protein, fat content) more accurately than other variants, even without direct training on this information. This suggests that understanding chemical relationships can act as a shortcut for the AI to grasp broader culinary concepts.

Multilingual and Cross-Cultural Capabilities

Epicure stands out with its multilingual training data, processing 4.14 million recipes from eleven sources in seven languages, including Chinese, Russian, Vietnamese, Turkish, Indonesian, and German. This is a significant leap from previous models like FlavorGraph, which relied primarily on English-language recipe corpuses. A sophisticated pipeline, utilizing Claude and Gemini embeddings, cleans and translates approximately 200,000 raw terms into 1,790 clean ingredients.

While the corpus significantly expands linguistic coverage, it remains unevenly distributed. Roughly half the material originates from East Asian sources, with Latin American, Eastern European, and South Asian cuisines each contributing single-digit percentages. This implies that while the model has a broader reach, its recommendations for South Asian ingredients might be less stable or comprehensive compared to East Asian or Western cuisines. For Indian users, this highlights the ongoing need for more region-specific data to fully leverage such tools.

Practical Applications for Indian Teams

For Indian F&B businesses, food tech startups, and even home chefs, Epicure offers several potential applications. A model that can distinguish between recipe companions and flavor relatives, translate ingredients across cuisines, or shift recommendations along axes like "fatty" or "fermented" could be invaluable.

For example, a restaurant chain in India could use Epicure for menu development, exploring novel ingredient combinations that resonate with local palates while maintaining flavor harmony. During supply chain disruptions, the model could suggest suitable replacements, helping maintain consistency. Its ability to shift a seed ingredient towards a target direction – e.g., turning "rice" slightly towards South Asia to yield curry leaf, urad dal, and fenugreek seeds – demonstrates its potential for cross-cultural culinary exploration and localization of recipes.

However, the uneven distribution of training data means that Indian teams should exercise caution and validate recommendations, especially for less represented regional Indian cuisines. The model's utility will likely increase as more diverse global culinary data becomes available and integrated.

Future Outlook and Availability

The underlying research and model weights for Epicure are now available on Hugging Face, allowing for independent verification by researchers and developers. A demo of an older version of the model is also accessible at epicure.kaikaku.ai.

Kaikaku.AI, founded in London in 2023, operates a robotic restaurant, Common Room, and has raised approximately $1.8 million in a pre-seed round in 2024. Their interest in a machine-readable map of the ingredient world aligns with their goal of automating and optimizing food preparation and inventory management. This development underscores the growing intersection of AI, food science, and automation, paving the way for more sophisticated culinary tools.

Source: The Decoder – https://the-decoder.com/ask-ai-what-goes-with-chicken-and-the-answer-depends-on-whether-it-learned-from-recipes-or-molecules/