Technology for voice AI
Twilio, managed voice APIs and your store integration.
Twilio Programmable Voice
The telephony layer supplies a number, outbound calls, routing and call events through an API. Twilio was used for calls in the earlier DooxSwitch project. For each store, we verify caller ID, permitted destinations and rates.
The conversation layer
The baseline pilot architecture uses Twilio ConversationRelay for speech recognition and synthesis, exchanging text with a dialogue server over WebSocket. An LLM API answers within the approved scenario and order data. Twilio documents a Ukrainian configuration; voice quality and comprehension are validated in test calls.
An alternative is ElevenLabs Agents with its native Twilio integration. The final option depends on measured dialogue quality and pilot cost. A working telephone connection does not by itself make an AI agent ready.
Integration service
An HTTPS API receives order events. A job queue manages calling windows and retries. Unique event keys prevent duplicate calls; signed webhooks return call state. ConversationRelay requires a separate service supporting persistent WebSocket connections, such as Node.js/TypeScript in a container, deployed separately from the website.
CRM and controls
JSON context contains the necessary order ID, customer contact and permitted terms. Outcomes such as confirmed, declined, callback_requested, needs_manager and no_answer map to your CRM workflow. Keys remain on the server. Payment credentials are excluded from dialogue context.
Grav CMS on Hostinger
Grav manages the website, Ukrainian and English articles and profiles. Gemini helps produce editorial content from verified sources. Order data and call audio are kept outside public CMS content.
Sources: Twilio ConversationRelay, languages and voices, ElevenLabs + Twilio.