CASE STUDY · OUTBOUND VOICE AI

Telefonia: AI outbound calls connected to Bitrix24

A delivered production system for recovering missed inbound leads: it selects eligible CRM records, calls through Asterisk, holds a controlled natural-language conversation, schedules consultations, and returns the complete outcome to Bitrix24.

Build something similar ↗View architecture

Anonymized client review for the Telefonia voice AI project — Selected project material · anonymized where required

335dialogue tests

3concurrent calls

5maximum attempts

2:55prototype recording

EARLY WORKING PROTOTYPE

Authentic project audio

The delivered version later improved pacing, intonation, interruption handling, dialogue stability, booking, and CRM synchronization.

THE CHALLENGE

What had to remain reliable.

The difficult part was coordinating streaming audio and unpredictable human dialogue. The system needed to distinguish caller speech from its own playback, recover from interruptions, separate a callback from a consultation booking, prevent duplicate CRM activity, and preserve an outcome after a partial failure.

THE SOLUTION

How the system was designed.

Critical intents are deterministic, while the LLM operates inside a constrained state machine. Call state persists independently of the live audio connection. Barge-in cancels synthesis and rejects self-echo, retry limits respect calling windows, and CRM checks prevent duplicate outreach.

DELIVERED

The working scope.

End-to-end lead queue, Asterisk AMI originate, SIP trunk, AudioSocket, and real-time two-way AI conversation.
Phone normalization, deduplication, funnel-stage checks, and open-activity checks before a call enters the queue.
A controlled dialogue state machine for greeting, discovery, consultation offer, booking, questionnaire, callback, and closing.
Russian scheduling logic for weekdays, relative dates, time ranges, occupied slots, business hours, and confirmation.
Outcome, manager note, extracted facts, transcript, stereo WAV, recording, and consultation activity synchronized to Bitrix24.

ARCHITECTURE

The main system boundaries and data flow.

[ Bitrix24 funnel / missed leads ]
                 | eligibility checks
                 v
       [ HTTP webhooks / call queue ]
                 |
       [ Dialer + retry policy ] ---> [ persistent call state ]
                 | AMI
                 v
      [ Asterisk / SIP ] <----------> [ customer phone ]
                 | AudioSocket
                 v
      [ VAD + streaming speech-to-text ]
                 v
 [ intent rules + state machine + LLM + RAG ]
                 v
      [ streaming TTS + barge-in ]

After call: facts + transcript + recording -> Bitrix24

OUTCOMES

What the architecture enables.

Missed leads returned to a managed calling workflow

Natural interruption without losing dialogue state

Conservative retry and concurrency controls

Complete CRM context after every completed or failed call

STACKPythonAsteriskAudioSocketDeepgramOpenAIElevenLabsBitrix24SQLiteDocker

YOUR PROJECT CAN START WITH A SHORT TECHNICAL REVIEW

Show me the workflow, integration points, and current bottleneck.

Discuss the project ↗