Multilingual Voice AI & Runtime Diagnostic Harness
AI receptionist for healthcare: voice agents for scheduling and clinic operations. Outbound call reliability, LiveKit dispatch, UAE data residency and the MENA launch.
Multilingual Voice AI & Runtime Diagnostic Harness
Enterprise healthcare launches demand low-latency multilingual voice over live telephony and immediate root-cause visibility.
Engineered end-to-end voice & harness reliability layers spanning Twilio/LiveKit ingress, ASR/LLM orchestration, deterministic evals, and trace telemetry.
- ▸ Removed ~15 s of dead air at pickup on outbound calls: the agent was greeting the caller while the line was still ringing.
- ▸ Made agent dispatch reliable on LiveKit: retry-and-verify, no duplicate agents in one call, no silent calls.
- ▸ Designed UAE audio-residency guardrails so TTS and STT fallbacks stay within approved regional providers.
At Sully the agent greeted callers while the line was still ringing, so they picked up to ~15 s of silence. The fix: wait until the person speaks, then greet. Here is that building block, running in your browser: Silero VAD v6, a 0.5 MB model that hears speech in 32 ms frames.
Audio stays on your device; nothing is recorded or sent.
Allow the microphone to start. Loading the model (0.5 MB): Measuring the room for a moment. Stay quiet… Listening. Say hello. Heard you after — ms. — ms of compute for that frame on this device. Didn't hear anything for 15 s: about as long as a Sully caller waited. Try again closer to the mic. Microphone access was blocked. Nothing was recorded. Allow it in the site settings to try. No microphone found. This browser can't run the demo: it needs microphone access and AudioWorklet. Paused to save your data: the demo downloads a 0.5 MB model. On a faster connection it runs right here. Something went wrong. Nothing was recorded.
Measured from the 32 ms frame where your voice rose above the room to the model’s decision, plus the compute time. Your voice can start anywhere in that first frame, so it may read up to 32 ms high; the look-back stops at about a second. Mic and OS buffering before the page gets the audio are not counted. At Sully the win was not a faster model but the agent waiting for a voice instead of a ringing line.