Mzia

A multilingual voice AI receptionist, with engineering focused on turn-taking, interruptions and the call lifecycle.

My role
I lead Mzia and am responsible for the whole architecture.
Mzia voice engineering: a caller interrupts, the relay manages the turn and checks for resumed speech
A conceptual view of call control: interruptions, session lifecycle and failure testing.

Product focus

Mzia handles spoken conversations in Georgian, Armenian, Russian and English. The engineering challenge is what happens between turns: a caller interrupts, the model pauses, a connection fails or someone hangs up while the session is still opening.

Engineering decisions

Measure the voice path

Twilio Media Streams connects to Gemini’s native audio through the relay. Model-comparison probes use the same audio and configuration; latency probes vary instruction size to examine its effect on response time. This makes provider and prompt choices testable.

Own the call lifecycle

The relay coordinates interruptions, bounded pre-session audio buffers, silence watchdogs and cleanup. A hang-up during connection closes the late-arriving model session; a provider disconnect ends the call instead of leaving a dead session open.

Keep audio independent of storage

Call metadata writes are asynchronous, so audio streaming does not wait for database completion. Call control and data capture have separate responsibilities within the relay.

Test failure paths

Integration tests run the relay with real HTTP/WebSocket connections and a controlled model double. Cases include interruption behaviour, failed connections, mid-call model drops, buffer overflow and hang-up races; live-model probes complement these offline checks.

Stack · Node.js / Express / WebSockets / Twilio / Gemini 3.1 Flash Live / Supabase / PostgreSQL

Architecture directionPlanned

Conversation and execution, separated

The next architecture centres on a persistent service request that can continue beyond one call. The voice layer handles dialogue; a domain service owns validation and business actions. Versioned requests discard stale lookup results after a correction, and idempotent actions prevent duplicate writes. This is a documented design direction; implementation is planned.

Recognition

1st place, Georgian Chamber of Commerce “AI x Business” hackathon — 18 June 2026.

Explore the product

Have a product or workflow in mind?

Tell me what it needs to do and where you are starting.

Email me about a project