Sales Voice & Calling Assistant
Duplex Speech-to-Speech & Outbound Telephony
Ultra-low-latency real-time voice agents handling 70% of inbound sales qualification and 25+ weekly candidate screening calls.
The Engineering Challenge
Sales qualification and initial candidate interviews were bottlenecked by scheduling latency and high candidate drop-off.
Core ProblemLegacy cascaded STT→LLM→TTS pipelines suffered ~3-second response latency, causing unnatural turn-taking and 40%+ candidate abandonment.
- Turn latency must feel conversational (< 400 ms)
- Support natural barge-in (interruption detection) without false audio cuts
- Reliable SIP telephony integration for outbound dialing over standard phone networks
System Architecture
End-to-end data pipelines and component isolation boundaries.
Duplex speech-to-speech architecture running on LiveKit WebRTC transport coupled to Gemini Live streaming APIs, with NumPy audio resampling over WebSockets and tuned Voice Activity Detection (VAD).
Key Engineering Decisions
Duplex S2S over Cascaded Pipeline
Replaced three-stage STT→LLM→TTS with native duplex speech-to-speech to eliminate audio serialization overhead.
Context Window Compression
Engineered conversation memory summarizer compressing 25k token histories into 10k token semantic checkpoints to maintain voice turnaround speed.
Silent Observer Agent via SSE
Decoupled analysis agent streaming live sentiment, recommended questions, and BANT lead scores to rep dashboard via Server-Sent Events without impacting audio pipeline.
Production Verification & Results
- Reduced conversational turn latency from 3,000 ms to 200 ms with natural barge-in
- Autonomously processes 70% of inbound client discovery conversations
- Executing 25+ live first-round recruitment screening calls every week over outbound SIP telephony
Production Stack & Tooling
Next Iterations & Roadmap
- On-device acoustic echo cancellation for noisy mobile microphone environments
- Real-time voice cloning adaptation for personalized recruiter matching
Want to build something similar?
I specialize in taking complex agentic workflows, duplex voice architectures, and high-throughput computer vision models into production.