LLMs are built by training a neural network on enormous datasets of text so it learns statistical patterns of language - grammar, meaning, and context - well enough to predict and generate coherent responses. Once trained, a base model is typically adapted for a specific use case through fine-tuning on domain-specific data, and often paired with a retrieval system that feeds it relevant, up-to-date information at the moment of a conversation, rather than relying solely on what it memorized during training.
Not all LLMs handle Arabic equally well, and even fewer handle Gulf or Levantine dialects, Arabizi, and Arabic-English code-switching with the same fluency they bring to English, since most base models are trained predominantly on English-language text. For an AI concierge operating across MENA banking, retail, and healthcare, the underlying LLM choice, plus how heavily it is fine-tuned on dialect-specific and industry-specific data, determines whether intent detection and responses actually feel natural to customers writing in their everyday spoken Arabic.
In Eshal: Eshal builds its conversational layer on large language models selected and fine-tuned for Arabic dialect performance rather than relying on an off-the-shelf, English-first model. This includes tuning for Gulf and Levantine Arabic, Arabizi, and code-switched messages, paired with retrieval of client-specific information such as order status or policy details, so responses stay accurate and grounded rather than relying only on general model knowledge.