Meydan Free Zone — AI Voice Agent
- <800ms
- Real UAE lines
- Barge-in
- CRM-scored
Highlights
- Outbound AI calls conduct structured lead qualification conversations
- Sub-800ms latency with barge-in handling for natural conversation
- Bridged LiveKit WebRTC agent rooms to real UAE phone lines
- Adopted Yeastar PBX for UAE carrier compliance and local DID numbers
- Full call transcription logging and structured lead scoring to CRM
Overview
I built a production AI voice agent for Meydan Free Zone, a UAE business-setup consultancy. The flow: a website visitor fills a lead-capture form (name, area of interest, business activity, country), and the system triggers an outbound AI phone call. The voice agent conducts a structured qualification conversation — capturing the lead’s details, asking qualifying follow-up questions, scoring lead intent, and logging everything (structured answers + score) back to the CRM for prioritised human sales follow-up.
This wasn’t a chatbot. It was real AI making real phone calls to real people in the UAE, and it had to feel natural.
The challenge
Conversational AI over telephony is brutally hard. The requirements:
- Real phone lines in the UAE — not a sandbox; real PSTN connectivity with carrier compliance.
- Natural conversation — sub-800ms latency, or callers hang up.
- Barge-in — the agent must react when the caller interrupts, like a real human.
- Structured output — the call must capture specific data and a score, not just a transcript.
- Reliability — appliance-grade telephony, because dropped calls mean lost leads.
What I built
Voice AI stack
LiveKit WebRTC agent rooms as the conversational backbone, bridged to real UAE phone lines. Azure Speech for STT/TTS and Azure OpenAI for LLM turn-taking with tool calling — so the agent can ask, capture, and confirm structured information during the call.
Telephony: why Yeastar
I evaluated FreeSWITCH and Asterisk for SIP/PSTN connectivity, but adopted Yeastar PBX in production for UAE-specific needs: local UAE DID numbers, appliance-grade reliability, and carrier compliance. The right tool for the regulatory and reliability context — a classic build-vs-buy decision.
Conversation quality
Sub-800ms end-to-end latency so the conversation feels live. Barge-in handling so the agent stops and listens when the caller speaks over it — the difference between “creepy robot” and “natural assistant.” Full call transcription logging for every call.
Structured lead capture → CRM
The agent doesn’t just talk — it captures: name, area of interest (company setup, licensing, banking), company/business area. It asks qualifying follow-up questions to score lead intent, then logs all structured answers plus the score back to the CRM. Sales gets a prioritised queue, not a raw transcript.
Evaluated alternatives
I evaluated Pipecat as a potential conversational framework during the build — the kind of option-weighing that’s normal when you’re on the edge of a new capability.
Why it matters
Voice AI in production, on real phone lines, with natural conversation and structured business output — is one of the hardest “AI product” problems to ship. This one ran in production for a paying client with real compliance constraints.
My role
Architect and builder — the telephony bridging, the voice AI pipeline (LiveKit + Azure Speech/LLM), the structured conversation design, and the CRM integration.