Latimal
A food AI research lab.
Generic embedding models treat "Murgh Makhani" and "Butter Chicken" as unrelated strings. They can't handle the transliterations, misspellings, and promo noise that litter real menus. They don't know that a "Classic Burger" and a "Veggie Burger" have a dietary conflict, or that "Kadhai Chicken" and "Karahi Chiken" are the same dish.
Food delivery needs its own AI. One that understands how people actually name, misspell, translate, and describe food across languages and cuisines. That's what we build.
The model
dish-embed is a fine-tuned bge-m3, 568M parameters, trained on 400K+ real menu item pairs sourced from production delivery platforms. It uses a two-stage retrieval pipeline: a fast bi-encoder for candidate generation, followed by a cross-encoder reranker for precision.
We evaluate against 18 internal benchmarks spanning Indian, global, cross-lingual, and category-specific scenarios. The model covers all major world cuisines, with explicit cross-lingual support for Japanese, Korean, Arabic, Spanish, Vietnamese, Thai, and Chinese menu items.
Team
Aditya Patni, founder.