Senda

Small clinics in Spain run their front desk on WhatsApp. They lose bookings for two reasons: nobody answers at night, and nobody follows up a few days later. Chatbots solve the first. They can’t solve the second, because good follow-up needs memory that lasts weeks, timing that feels human, consent you can prove, and cost discipline. Every business-initiated WhatsApp message outside the 24-hour service window is a paid template. Senda is an agent that lives in a clinic’s WhatsApp and handles each lead end to end: Understands real customer input. It handles text, noisy voice notes (transcribed with Muse Voice Transcribe) and photos (Muse Spark vision), in peninsular Spanish. Books real appointments. It checks actual availability against the clinic’s opening hours and Google Calendar, books the slot and sends a deposit payment link. Remembers. It stores structured facts (name, payday, main objection, availability) plus semantic recall over the whole conversation with pgvector. If a patient says “I get paid on the 15th”, Senda schedules the follow-up for that date. Follows up like a person. It runs declarative follow-up sequences with quiet hours (09:00–21:00 Madrid time, no Sundays), a per-contact cap and a minimum gap between messages. Any reply from the customer cancels pending sales follow-ups automatically. Knows what each message costs. It writes free-form messages inside the 24-hour window, switches to approved templates outside it, prices every send by category (marketing / utility), and reports savings against a fixed-cadence bot. Hands off cleanly. Urgent symptoms, complaints, clinical questions or a request for a human route the conversation to staff with a summary, and automated follow-up pauses. GDPR by construction, not by prompt Consent is handled by a deterministic layer that runs before the LLM. The model never decides on consent. The layer recognises SÍ / NO / BAJA / ALTA, and every consent event is stored with the customer’s literal words, timestamp, purpose, legal basis and policy version (Art. 7.1). BAJA cancels all scheduled messages and wipes follow-up memory. The memory layer refuses health data (Art. 9). The system also includes data export (Art. 15/20), erasure (Art. 17) and a 30-day retention purge.