Building Velyn Dental: First Production Tests
Early Results from Production Testing
We've been running Velyn Dental in limited production for two weeks now with a handful of beta practices. The system is handling real patient phone calls, and we're learning a lot.
So far, response times are holding steady around 2 seconds (caller speaks → AI responds), which feels natural in conversation. We've had a few hiccups during peak hours, but nothing critical.
The Technical Stack
Here's what we're using to keep things running:
Infrastructure:
- Vercel Edge Functions for hosting the API endpoints
- Upstash Redis for basic rate limiting
- Supabase Postgres for storing call data
- Twilio Voice for the actual phone service
- Datadog + Sentry for monitoring and error tracking
Basic Guardrails:
- Rate limiting to prevent abuse (10 calls/min per IP)
- Webhook signature verification from Twilio
- Simple honeypot fields to catch spam bots
- Error handling for when upstream services are slow
Monitoring:
- Tracking response latency (aiming for under 2s)
- Error rate monitoring (we alert when it spikes above 1%)
- Basic call volume dashboards
- Sentry for catching exceptions
What We're Learning
Two seconds feels fast enough. With streaming responses and edge functions, we can make real-time AI conversations feel natural.
Monitoring is critical. When something breaks at 2am, we need to know immediately. Our Datadog setup is already paying dividends.
Building call coverage for real practices is harder than demos. Every edge case matters when it's a real customer on the phone.
What's Next
We're onboarding a few more beta customers this month to stress-test the system. Need to see how it handles higher call volumes before we open it up more broadly.
The goal is simple: build something reliable enough that dental practices can actually depend on it.
About Autonomy AI®
Autonomy AI® builds Velyn Dental for dental call coverage and BuckHound™ for price tracking. Learn more at auai.cloud.