Sawt
AI receptionist for MENA businesses — calls and chats in local Arabic
- 5
- channels, one agent
- 1.4s
- time to answer (from 3.7s)
- 600+
- commits shipped
- 3
- UI languages (RTL)
Hear the pipeline work
A scripted clinic booking, replayed in four dialects. Each stage lights up as audio, transcripts and tool calls move through it.
// what's running right now
- 01
Caller
Phone via the business's own PBX (SIP), or the website widget
- 02
Twilio Media Streams
8 kHz μ-law audio over a WebSocket bridge
- 03
Gemini Live
Speech to speech, locked to the caller's language
- 04
Agent tools
Calendar, RAG over documents, leads, transfer
time to answer · Gemini Live
- 5
- channels, one agent
- 1.4s
- time to answer (from 3.7s)
- 600+
- commits shipped
- 3
- UI languages (RTL)
Stack
Overview
Sawt (صوت — "voice") is a production, multi-tenant SaaS for small businesses in the Arabic-speaking world — clinics, salons, restaurants and shops in Jordan first, then Saudi Arabia and the UAE. An owner builds an agent in minutes from an industry template, and it answers every channel at once: a one-script website widget (voice, chat or both), WhatsApp, Instagram and Messenger (including photos and voice notes, which it can answer with a voice note), and phone over the business's own PBX via SIP — no foreign number to buy. Calls run speech-to-speech on Gemini Live over a WebSocket bridge from Twilio Media Streams or the browser; I cut time-to-answer from 3.7 s to 1.4 s, and premium agents speak in ElevenLabs voices. Chats run on Gemini 2.5 Flash with the same tools: a pgvector RAG knowledge base, calendar booking that refuses taken slots, lead capture, payment links, and transfer to a staff member who presses 1 to accept. Owners get one inbox for every call and chat with transcripts, recordings, topic and sentiment, can watch a live call and take it over, test an agent against simulated callers, restore any saved version, and receive a weekly email of missed chances and unhappy customers. Sawt is also an MCP server and can use a customer's own MCP tools. Arabic, English and French UI with full RTL, Dodo Payments billing, row-level security on every table enforced in CI, and near-zero idle infrastructure on Vercel and a suspend-when-idle Fly.io machine.
Highlights
- 01
Five channels from one agent: website widget (voice + chat), WhatsApp, Instagram, Messenger and phone over SIP from the business's own PBX
- 02
Speech-to-speech calls on Gemini Live over a WebSocket audio bridge — time-to-answer cut from 3.7 s to 1.4 s; premium agents speak in ElevenLabs voices
- 03
Local Arabic dialects with Arabic, English and French UI in full RTL; answers photos and voice notes, and can reply with a voice note
- 04
Real tools, not promises: RAG over uploaded documents, calendar booking that refuses taken slots, leads, payment links, and human transfer with press-1 confirm — replies that claim an action nothing did are held back
- 05
Owner console: unified inbox, call recordings with consent, topic + sentiment labels, live listen-in and takeover, simulated-caller testing, agent version history, weekly insights email
- 06
Production SaaS: workspaces and roles, MCP server + customer MCP tools, Dodo Payments billing, RLS on every table enforced in CI, near-$0 idle infra
Want to talk about how this was built?