All work

Case 01● Beta2026 — now

VoiceManager

An AI outbound sales rep: a Russian-speaking voice agent

Role
Creator — from architecture to production
Period
2026 — now
Stack
Python · LiveKit · Yandex SpeechKit · Gemini · ElevenLabs · Silero VAD · FastAPI · PostgreSQL · Docker
VoiceManager dashboard: pick-up rate, orders, call funnel and activity

The problem

Outbound phone sales are expensive and hard to scale. VoiceManager takes over the first line: the agent dials through a contact list, holds a natural conversation and hands a ready order to the CRM — items, address and preferred delivery time included.

How it works

  1. 01audio

    Call

    SIP / WebRTC

    outbound call from a contact list

  2. 02audio

    Media server

    LiveKit · self-hosted

    Silero VAD and a Russian turn detector

  3. 03audio → text

    Speech-to-text

    Yandex SpeechKit v3

    directly from Russia

  4. 04text

    Agent reply

    LLM adapter

    V2Horizon gateway → Gemini, YandexGPT or OpenRouter

  5. 05text

    Normalisation

    Normalizer

    numbers and stress marks before synthesis

  6. 06text → audio

    Text-to-speech

    ElevenLabs

    Flash v2.5 or v3 + on-disk phrase cache

Audio and recordings stay in RussiaExternal APIs see text only↺ The reply streams back into the call via LiveKit

The voice pipeline runs on self-hosted LiveKit: Silero VAD and a Russian turn detector tell when the person has finished speaking. Speech is recognised by Yandex SpeechKit v3 directly from Russia. Replies come from an LLM behind a pluggable adapter — by default my own V2Horizon gateway to Gemini, with YandexGPT direct or OpenRouter via proxy as alternatives. Text is normalised before synthesis (numbers, stress marks) and voiced by ElevenLabs — Flash v2.5 by default or v3 for quality — with an on-disk cache for frequent phrases.

Client dashboard

  • Moderated campaigns: greeting, product facts, objection handling, upsells
  • Contact lists, dialing and a test environment to hear the agent right in the browser
  • Call logs: transcript, outcome (order, call back, refusal), duration and CRM-ready data
  • Dashboard: pick-up rate, order conversion, call funnel and 30-day activity
  • Security: two-factor sign-in, integration API keys, session management
  • The AI assistant answers questions, pauses campaigns and edits settings — every action is logged in the chat

Screens

5