Vocalendaanswers the phone,so they don't have to.
Multi-tenant voice booking with Google Calendar + SMS.
The phone still rings. Nobody picks up.
The founders came in with a working prototype and a problem: their AI voice agent sounded plausible in a demo, but the product around it — onboarding, dashboards, billing, voice configuration — was a Frankenstein of three different no-code tools. They had paying customers and a churn rate that wouldn't stop bleeding.
My job was to compress the surface area into a single product that a salon owner could set up in twenty minutes, and to give the brand the gravity it needed to charge enterprise prices to a non-enterprise audience.
Twenty minutes from sign-up to a live agent taking real calls.Project north star · Q1 2026
Sell the outcome, not the model.
Most voice-AI products lead with the technology — model names, latency numbers, the word "agentic" somewhere above the fold. The audience for Vocalenda doesn't care. They care that Tuesday's missed call is now booked into next Wednesday at 2pm with a confirmation text already sent.
We rewrote the entire positioning around the receptionist who never sleeps — a framing that lets the product compete with hiring decisions instead of with other AI tools. The visual language followed: confident, restrained, slightly old-fashioned in the way independent professionals are.
One product, three surfaces, one feel.
The brand had to hold up across a public site (where the audience is shopping), a dashboard (where they live for five minutes a day), and the voice itself (which they never see but always hear). A single typographic system, one signal colour, and a tight motion vocabulary tied the three together.
The voice took the longest. We A/B-tested seven different agent personas with real customers before settling on the one that booked the most appointments without sounding like a customer-service script.
Multi-tenant, typed end-to-end.
Every tenant lives in its own Postgres row-level-security boundary. A new tenant provisions a Twilio number, a Google Calendar OAuth, and a tuned agent config in under a minute. The whole pipeline is typed from the Postgres schema up to the React props, so a missing field on a tenant config fails at build time, not at 9am on a Monday.
- Tenants live
- 184
- Avg setup time
- 17 min
- Missed-call rate (before)
- 34%
- Missed-call rate (after)
- 1.8%
- Bookings handled / mo
- 11,402
- Churn (post-launch)
- 3.1%
Radu shipped what would have been a six-month, three-person engagement in fourteen weeks alone. He's the only contractor I've ever hired who could argue me out of a feature.