ProntoVA

The comfortable version of this is a general-purpose WhatsApp chatbot framework. I picked a buyer instead: appointment-based clinics in Madrid — dental, aesthetic, physio. They take inquiries on WhatsApp all day, the owner answers personally, and because they are health-adjacent, Spanish data-protection guidance turns GDPR from a tax into the differentiator. Then I built the distribution alongside the software: a researched target list in a self-hosted CRM, a written outreach playbook, and a pilot offer where the client sets the success criteria.

Choosing a real buyer is what killed my channel, and that is the decision worth reading. V1 ran on Baileys, an unofficial WhatsApp library: QR login, no Meta approval, live in a day — and it worked, a real five-turn conversation through the full pipeline on 2026-07-07, ten messages persisted and metered end to end at $0.026. Then I deployed to a VPS and it failed deterministically: WhatsApp accepted the pairing, registered the device, then killed the authenticated session five seconds later with a 401. Residential connection fine, datacentre never. IP policy, not a bug I could fix.

A workaround existed — residential egress. I didn't take it. The bar is not "can I link once for a demo", it is "can a clinic depend on this for weeks without their business number being banned", and that channel had already silently de-registered my own working session inside an hour. So on 2026-07-08 I demoted Baileys to zero-stakes internal demos and made the official WhatsApp Cloud API the pilot channel: a bigger build, and one that has to clear Meta's business verification first. I traded my launch date for a channel a paying client can depend on.

Live at prontova.net. The runtime — app plus its own Redis, against Postgres over a private network — has run 24/7 on a VPS since 2026-07-08. Channel, model, payment and speech-to-text are each swappable per project in config, which is why the channel decision above is a new adapter, not a rewrite. 194 automated tests pass on the gate every goal had to clear — one further test, a live-API smoke, is opt-in and skipped by default. Tenancy is enforced by isolation tests, every model call is metered to a cost, and the privacy layer — retention purge, erasure, opt-out, PII redaction in logs — shipped and was reviewed before any clinic could be onboarded.

One review catch is a product judgment, not a code bug: my opt-out keyword list originally matched the bare Spanish words cancelar and baja. In a dental clinic those mean "cancel my appointment" and "sick note" — a patient typing either would have been silently and permanently blocked. Caught before it shipped; the list narrowed to unambiguous unsubscribe phrases, and the test now asserts those two inputs must not match.

Repo link placeholder — private today.