VoDialer AI: Real reply gap of ~1.2 s on live PSTN calls
AI outbound voice dialer built as a full VICIdial replacement for US Medicare call centres.
The calls that shaped it
Replace the whole stack, not just the agent
Most 'AI calling' products sit on top of a legacy dialer. VoDialer rebuilds the dialer itself: eight Rust services on FreeSWITCH and Kamailio, with SignalWire for PSTN. The AI agent is a first-class caller, not a plug-in.
Latency is the product
The real-time cascade — AssemblyAI speech-to-text, DeepSeek as the LLM, ElevenLabs or Fish for text-to-speech — is tuned for a ~1.2 s reply gap on live calls. Every architectural choice (Rust, streaming, sentence-level TTS) serves that number.
Operate it like a regulated system
A platform console with roles, TOTP and audit logs, plus an Expo mobile app. US Medicare call centres are a compliance-heavy market; the controls were designed in, not added after launch.
What this proves about how I lead
VoDialer is the hardest system I have built alone: telephony, real-time audio and LLM orchestration in a memory-safe language. It proves I can take an engineering team into unfamiliar, low-level territory and still ship — and that I measure success in production numbers (reply gap, concurrency) rather than demos.
Every product under Atrix was built with the same operating model I bring to a team: a written architecture decision before the first commit, a thin vertical slice in production within the first weeks, AI-assisted delivery for the volume work, and human judgment reserved for the decisions that compound — data model, security boundary, provider choice. The hire-me page lays out how that translates into a 90-day plan for your company.
Want to talk through the architecture?
I'm happy to walk a hiring panel or a founder through any of these systems in detail — or ask my AI twin first.