VoiceCare Scheduler — voice clinic booking

Elderly patients often depend on family to book clinic visits through apps with tiny text. One confusing screen and they give up — or travel to the clinic just to ask. VoiceCare Scheduler lets them speak instead, on the AssemblyAI Voice Agent API (agents.assemblyai.com: short-lived token minted server-side, live browser voice WS, English inline session with keyterms + barge-in turn-taking, vendor TTS voice, 2 tools check_availability/book called as tool.call->tool.result): say the need, the agent transcribes, extracts day/time entities, asks for confirmation aloud, then books only after the patient says YES (strict, transcript-proof). STT batch/WS fallback (Universal-3-5-pro, en, diarization) keeps the demo runnable. If the user interrupts mid-flow, the agent stops, listens, and continues with the better-understood sentence. If confidence is low, it never guesses — it escalates to a human hotline. Every run prints per-stage latency (reproducible locally). Measured live: WS handshake ~1.0s, PARTIAL ~1.2s, FINAL ~3.9s (Speaker A, `sample_en.wav`); batch ~8.5s with entities. Accessibility-first: slow clear TTS, large-text cards and autoplay replay serve elderly users, while barge-in turn-taking and never-guess escalation keep patients with unclear pronunciation safe instead of misbooked. Users: clinics serving elderly patients who don't use apps. Market: ~12M over-60 VN (UNFPA 2023) + ~5k private clinics (MOH est.); SAM 500 x $39/mo = $234k ARR yr1. Revenue: per-clinic SaaS $39/$99 per month, cheaper than a $250 phone receptionist. Track: Main Voice Agent (full operating loop: token + live WS + TTS + 2 tools + barge-in + YES-gated consent + escalation + measured latency; doubles as Accessibility entry for elderly/low-clarity speech). Live demo + video with EN overlays showing the full loop.