All projects
case study2026·live

Sawt

Arabic AI voice agents for MENA — real phone calls, real-time

Stack

Next.js 16FastAPISupabase (pgvector)TwilioOpenAI RealtimeDeepgramElevenLabsFly.ioVercel

Overview

Sawt is a production, multi-tenant voice-AI platform built for the MENA market: AI agents hold real phone conversations in Arabic dialects to take bookings, answer questions, capture leads, and hand off to humans. The core is a real-time media bridge that streams live call audio through Twilio, Deepgram speech-to-text, a function-calling LLM, and ElevenLabs speech synthesis — engineered to cut cost per call-minute by roughly 57% versus the OpenAI Realtime API while staying low-latency and natural. On top sits the full SaaS: 14+ agent tools and integrations (calendars, CRM, WhatsApp, Stripe payments), a RAG knowledge base, organization and role-based access control, usage-based billing (Paddle / HyperPay), and automated push-to-deploy CI/CD across Vercel and Fly.io.

Highlights

  • 01

    Real-time voice pipeline: live call audio streamed through Twilio, Deepgram STT, a function-calling LLM, and ElevenLabs TTS

  • 02

    ~57% lower cost per call-minute versus the OpenAI Realtime API

  • 03

    Arabic dialect support — Levantine, Khaleeji, Egyptian, and MSA

  • 04

    14+ agent tools and integrations: calendars, CRM, WhatsApp, Stripe payments, plus a RAG knowledge base

  • 05

    Multi-tenant org / role-based access, usage-based billing (Paddle / HyperPay), push-to-deploy CI/CD on Vercel + Fly.io

Want to talk about how this was built?