The lights are off. Press the power button to turn them on and read this project.
An AI WhatsApp receptionist for clinics in Latin America, evolving into a two-sided healthcare and wellness platform. Agentic AI wrapped in safety and process layers dense enough to deploy in healthcare.
- Started
- 2025
- Status
- Active
- Stack
- Next.js 15, Supabase, Twilio WhatsApp
- Links
- ·
Mozt's Vita started as an AI WhatsApp receptionist for medical clinics in Guatemala (the Mozt Clinics product, currently pilot-ready) and is expanding into a two-sided healthcare and wellness platform for Latin America. The architectural bet under both surfaces is the same: keep the agentic AI loop, but wrap it in process flow and layered guardrails dense enough to deploy in healthcare.
The bet
Most "AI for SMBs" products ship a chat interface wired to an agentic LLM loop and let the model figure things out. The architecture is "ask the model, hope for the best." For a wellness or healthcare context that doesn't survive contact with reality. Patients ask sensitive questions. The model hallucinates a price. The owner gets paged on a Sunday because a kid mentioned an allergy. The agentic loop is powerful, but on its own it has no idea when to refuse, when to escalate, or when its confidence is low.
Our bet was to keep the agentic AI doing the agentic work, and surround it with layers of safety and process flow that turn an "ask the model, hope for the best" system into one that knows what it's doing. Inputs are screened before the agent runs. The agent operates inside a constrained workspace where it can only act on verified clinic configuration. Sensitive content classes route to humans rather than to the agent. Every decision the system makes is recorded for operator review.
Once the agent operates inside that envelope, we can tell a clinic owner what their bot will and won't do. That sentence is what makes a healthcare operator willing to give the bot a real WhatsApp number.
Agentic AI inside a process flow with guardrails. The model does the work. The structure around it keeps the work safe.
Patient sends a WhatsApp message. The system verifies the request, identifies the clinic, and loads the clinic-specific context the agent will work inside.
Sensitive content classes (medical questions, complaints, human-handoff requests, conversations involving minors) get identified and routed to a human before the agent runs. The agent never has to decide whether something is safe.
The agent takes the request inside a bounded workspace. It can only act on the clinic's verified configuration. It does not invent prices, hours, or policies; it does not freelance outside its scope.
The system schedules, answers, escalates, or captures the message as a lead. Each action class has its own safety profile and its own response treatment.
Every decision the system made gets recorded as a structured row. Operators see real-time transcripts, can pause or override at any point, and receive instant DMs on every escalation.
How it works
The clinics product started from a simple commercial observation. Clinics in Latin America run patient acquisition and retention on WhatsApp, and they lose money at four specific defects. Inbound latency (the receptionist can't reply at 9pm on a Sunday). Booking back-and-forth (six to eight messages to schedule one cita). No-show drift (twenty to thirty percent of scheduled revenue walks). Zero recovery on missed appointments (a no-show is a permanent lost patient). We didn't need a smarter chat product. We needed to remove the human-attention bottleneck on inbound and replace it with a process that could run twenty-four hours a day in the clinic's own tone.
What we shipped first was the engine: an agentic AI bot wrapped in the safety and process layers above. Multi-tenant from day one with row-level security on every table. Every patient turn produces a structured audit record that captures what the system saw, what it decided, what it did, and how confident it was. The audit is queryable so a clinic owner can answer questions like "how many escalations went to a human last week" without having to ask us.
The cockpit came next. The clinic owner doesn't run the bot, they supervise it. Real-time transcript view, manual reply, pause-bot per patient or clinic-wide, manual booking, owner DM on every escalation with a subject line that differentiates medical emergency, minor consent, complaint, and human request. The lesson from earlier work was sharp: never draw a button that doesn't have a handler. Every UI element maps to a real action.
That whole stack cleared the pilot-readiness bar in mid-April. Tests passing, type errors zero, audit rows queryable, escalations going to the right place. What changed after that wasn't the engineering, it was the product. The clinics work was always supposed to be the beachhead, not the company. Mozt's longer thesis is that the same pattern (agentic AI plus process flow plus operator cockpit) transplants to other SMB workflows. So the next product, Vita, took the engine and reshaped the surface around it.
Vita is a two-sided platform. A patient-facing PWA where people in Latin America can find a fisio, a psicólogo, a dentist, a nutricionista, book in three steps, and chat in-app. A provider-facing web tool where the practitioner runs their day with calendar, clinical notes, services, patient list, and team management for clinics that have more than one practitioner. And an internal operations console for the team supporting the platform.
The agentic AI lives inside Vita's provider surface, wrapped in the same kind of safety and process flow that runs the clinics product. Repetitive workflows that today require a receptionist's attention (confirmation chains, reschedule coordination, no-show recovery, follow-up loops) get routed through that process. The agent does the work. The structure around it keeps the work inside the lines.
We are deliberately not sharing pricing, specific clinic names, partner relationships, internal architecture, or the operator console. Those parts of the product live in pilot-grade discretion until we are ready to talk about them publicly.
What broke through
The shift wasn't a single insight. It was the realization that the value isn't in choosing between agentic AI and structure. It's in combining them. The agentic loop is what lets the bot handle natural language conversation, ambiguous patient requests, multi-turn context. The structure around it is what makes those conversations safe in healthcare. Without the structure, the agent is too unpredictable to deploy. Without the agent, the structure is too rigid to feel modern.
Once we settled into that combination, the work got cumulative. Every clinic-specific configuration could be added without changing how the agent works. Every new safety rule could be added without changing how decisions get audited. Every new vertical Mozt expands into can reuse the same layered stack with a different surface and a different config on top.
What I learned
The default engineering choice when shipping AI products is to lean entirely on the model. The harder choice is to keep the model where it earns its weight, surround it with structure where it doesn't, and accept that the resulting system is six months of work, not a Friday afternoon. The IE move was to treat the agentic AI as a single component inside a larger process, with its own quality gate and its own failure mode. Operators trust systems they can supervise. Models alone are not supervisable. Models inside a process are.