What we built

Cadence (cadenceagents.ai) is an AI voice receptionist for service businesses — dentists, vets, salons, plumbers. When the owner can't pick up the phone, Cadence answers, qualifies the lead, books the appointment, and texts a confirmation. It runs on Retell (the agent), Twilio (the phone numbers), and Google/Microsoft Calendar (the busy-check).

The shape of what we shipped

The full onboarding stack lives in PRs #229–#237. By the end of a 5-minute setup the customer has:

• A managed Twilio number bought and provisioned via Retell.
• A Retell agent trained on their business hours, services, and pricing.
• Calendar sync (Google or Microsoft) with real-time busy-check.
• An SMS confirmation pipeline.
• A trial that extends on the first real call (so they're not paying for a dead line).

Why "speed-to-lead" is the only metric that matters

If a service business misses a call, the lead is gone in 5 minutes. Not 5 hours — 5 minutes. The competing business that picks up first gets the booking. That's the whole product: be the first to pick up, always.

The technical bar is: a missed call must trigger a callback or a text within 30 seconds. Anything slower = lost lead.

The honest result

Cadence solves a real problem for service businesses. The hard part isn't the AI — Retell handles that well. The hard part is the integration sprawl: Twilio, Retell, Google Calendar, Microsoft Graph, Stripe, the customer's PMS (practice management system). Each integration has its own quotas, webhook flakiness, and edge cases.

What the next lessons cover

• Lesson 2: the speed-to-lead architecture — the webhook chain, the calendar busy-check, the fallback when Retell is down.
• Lesson 3: the production code — webhook handler + outbound dial + busy-check + the gotchas around Twilio rate limits.