Echo
Conducts natural phone conversations and books medical appointments live, where low-latency barge-in matters more than raw model scale.
Engineering highlights
The core architectural breakthroughs.
High-scale engineering demand distilled into four verified production milestones.
Sub-700ms voice turnaround creates natural human cadence
Stops synthetic audio immediately when patient interrupts or coughs
7.5x expansion in simultaneous clinical intake handling capacity
Zero-drop direct EHR scheduling verified over call readback
Executive overview
The engineering challenge & solution.
Patient appointment intake lines were overwhelmed, dropping 35% of inbound calls during morning peaks. Existing voice bots suffered from robotic 2.4-second response pauses and could not handle natural human conversational interruptions: if a patient coughed or said 'Wait, actually...', the bot talked over them and crashed the booking flow.
We engineered a custom WebRTC audio pipeline connecting SIP telephony to a low-latency streaming pipeline. Audio chunks are streamed to an incremental speech-to-text engine; dialogue state management executes speculative token generation; and a streaming TTS engine begins audio synthesis after the first 3 tokens.
Engineering governance
The three non-negotiables.
These are the architectural constraints that shaped every boundary — no trade-offs or compromises permitted.
Sub-700ms Total Voice Turnaround
From patient speech end to initial synthetic voice phoneme output must clock under 700ms over telephony.
Sub-50ms Instant Barge-in
When patient audio is detected mid-response, the voice synthesizer must stop immediately with zero audio echo.
Direct EHR Appointment Booking
Integration with 6 healthcare scheduling systems with strict patient verification and slot conflict locking.
Speculative Token Generation and Barge-in Audio Cancellation Over Telephony Jitter
Telephony packet jitter and background ambient noise frequently triggered false barge-in interrupts, causing the agent to stutter and abort legitimate sentences.
We implemented a dual-stage Voice Activity Detector (VAD) coupled with an acoustic echo canceller. The pipeline buffers incoming audio and evaluates phoneme probability before signaling an interrupt. In parallel, the dialogue engine speculates likely patient answers while they are speaking, priming the LLM context window ahead of time.
Request lifecycle
How data moves through Echo.
Production verification
Measured outcomes in production.
Every metric below was captured under real production traffic and audited against historical baseline data.
Operational resilience
Automated safety guardrails.
Systems fail in production. Here is how Echo survives network partitions, upstream timeouts, and anomalous inputs without human intervention.
Human Clinical Triage Escalation
Keywords like 'chest pain' or 'bleeding' immediately transfer to an ER nurse.
Redundant SIP Trunk Failover
Dual telephony carriers ensure call continuity during network disruptions.
Patient Confirmation Readback
Dates, times, and doctor names read back and confirmed before booking write.
Next step
Building a system with this level of demand?
Bring us the constraint that keeps your engineering leadership up at night — the latency ceiling, the compliance perimeter, or the unyielding reliability requirement.