Our Dubai voice agent projects are designed for polyglot call flows from the ground up. The architecture uses a language detection layer at the front: the caller's first 2-3 seconds of speech trigger a language classifier that identifies the spoken language and routes to the appropriate ASR (Automatic Speech Recognition) pipeline. We maintain parallel ASR models for Arabic (Gulf dialect), English, Hindi, Urdu, Russian, and Mandarin — the top 6 languages in Dubai customer service. For each language, the NLU layer classifies intent and extracts entities using language-specific fine-tuned models. Critical design decision: we do NOT translate everything to English for processing. Each language pipeline runs natively, preserving the nuance and context that translation destroys (a Hindi-speaking caller's complaint carries different cultural expectations than an English-speaking caller's). For Dubai government clients, we integrate with the unified government CRM and service platforms. For airlines, we connect to Amadeus, Sabre, or the carrier's PSS (Passenger Service System) for real-time flight data, booking modifications, and loyalty account access. For hotels, PMS integration (Opera, Protel) enables reservation lookups and guest service processing. TTS output uses language-matched neural voices — callers hear responses in their own language from voices that sound natural, not robotic. The entire system runs on UAE infrastructure (Azure UAE North) to satisfy NESA and TDRA requirements.