BlogCase Studies

Case study: MediVoz — an AI receptionist for Mexican clinics

Why a market too small for US healthtech — private clinics in Guadalajara and Monterrey — is exactly the shape of opportunity AI-native solo builders should want.

Private clinics in Mexico have a problem US healthtech won't solve: the receptionist leaves at 6pm, patients text WhatsApp at 9pm, and nobody answers until the next morning. The clinic loses the appointment.

MediVoz picks up that WhatsApp thread, books the appointment, syncs to the clinic's calendar, and sends a reminder the day before. In Spanish, with the regional courtesies. No human involved unless something unusual happens.

Why US healthtech can't serve this market

Rectangular economics. A typical US healthtech sale is $8–15 per patient per month, contracted annually, routed through a Zocdoc or SimplePractice sales team. That motion doesn't make sense for a 2-doctor clinic in Tepic charging 600 pesos per consult.

The actual ACV we can hit is ~$40/month per clinic, pay-as-you-go, WhatsApp-first onboarding. No US cap-table company is going to build that stack.

The architecture

  • Voice layer: Twilio SIP trunk into a Deepgram streaming ASR, Claude Sonnet for intent + response, ElevenLabs for TTS. Median latency around 700ms with barge-in detection so the agent yields when the patient starts talking.
  • WhatsApp layer: Twilio's WhatsApp Business API. Same Claude prompt, different modality.
  • Calendar layer: Supabase Postgres with RLS scoped per clinic, integrated with Google Calendar so doctors see appointments in whatever tool they already use.
  • Ops: Sentry + PostHog for per-clinic observability, BullMQ for appointment reminders.

What took the longest

Not the voice stack. Not the prompt. The hard work was the cultural part — training Claude to handle the specific formalities Mexican patients expect ("Con mucho gusto", "Para servirle"), to recognize when a patient is in distress vs just confused, and to know when to escalate to a human operator.

We spent two weeks labeling conversations. The voice pipeline took three days.

What I'd tell another solo builder

Find the market where the product is "technology US companies already have, but adapted for a context US companies don't understand." That's not one-off work. That's every Spanish-speaking country, every language, every underserved vertical.

It requires you to care about the niche — to have family in Guadalajara, to text in Spanish slang, to know why a clinic that takes solo efectivo can't use Stripe. Taste is a moat. So is location.