Kaikaku

https://t.co/n9WPl8YIhV Trains Food Models on 4.14 Million Recipes\n\nLondon startup https://t.co/n9WPl8YIhV released Epicure a set of three AI models that separate recipe pairing from chemical flavor similarity. The models were trained on 4.14 million recipes in seven languages plus the FlavorDB chemistry database.\n\nThe Cooc model learns ingredient co-occurrence in recipes while Chem learns shared flavor molecules and Core blends both. \n\nAsked about chicken Cooc returns garlic onion and black pepper while Chem returns beef and pork. \n\nFor basil Cooc suggests parsley olive oil and parmesan while Chem suggests oregano tarragon and rosemary. \n\nThe chemistry-based Chem model classifies taste and nutritional properties more accurately despite never receiving that data directly. \n\nA cleanup pipeline using Claude and Gemini embeddings reduced 200000 raw terms into 1790 ingredients with only about a third anchored in the chemical database. \n\nKaikaku raised about $1.8 million in a 2024 pre-seed round and released model weights and datasets on Hugging Face.