From Idea to Production in 56 Days
Loquent (getloquent.com) started in 2025 with a simple observation: healthcare clinics miss 30-40% of their incoming phone calls. Every missed call is a lost appointment, lost revenue, and a patient who books elsewhere. We set out to build an AI voice agent that could answer every call, book appointments, and handle patient inquiries — all with natural-sounding conversation.
Eight weeks later, the first version of Loquent was in production, answering real calls with sub-800ms voice latency. In August 2026 we rebuilt it as an AI workforce for small businesses, but the engineering from those first eight weeks still matters.
This is the story of how we built it.
The Problem We Solved
Canadian dental clinics face a persistent operational challenge. Front desk staff are pulled in every direction — checking in patients, handling insurance, answering phones. During peak hours, calls go unanswered. After hours, they go to voicemail. Most patients don't leave voicemails — they call the next clinic on the list.
We talked to clinic owners who estimated they were losing $5,000–$15,000 per month in missed appointments from unanswered calls alone. According to industry data, the average dental clinic misses 30–40% of incoming calls during peak hours, and fewer than 20% of patients leave voicemails.
The Tech Stack
We chose our stack for speed to production and reliability at scale:
- Voice Infrastructure: Twilio for telephony — handling inbound and outbound calls with reliable carrier-grade infrastructure
- AI Models: OpenAI GPT-4o for conversation understanding and generation, with Anthropic Claude as a fallback for specific reasoning tasks
- Backend: NestJS on Node.js — our standard backend framework that lets us move fast with strong typing and modular architecture
- Database: PostgreSQL on AWS RDS with Prisma ORM for type-safe database access
- Real-time Processing: Custom audio streaming pipeline for low-latency voice interactions
- Deployment: AWS infrastructure in the Canadian region for data residency compliance
Architecture Decisions That Mattered
Conversation State Machine
The biggest technical challenge was managing conversation state. A patient calling to book an appointment needs to navigate a multi-step flow: identify themselves, state their need, check availability, confirm a time, and provide contact details. The AI needs to handle interruptions, corrections, and tangents gracefully.
We built a state machine that tracks conversation context across turns while allowing the AI to deviate from the script when the patient's needs require it. This was the difference between a rigid IVR system and a genuinely useful voice agent.
Latency Optimization
Voice conversations are unforgiving — anything over 800ms of latency feels awkward. We optimized every step of the pipeline: pre-warming model connections, streaming audio in chunks rather than waiting for complete utterances, and running inference close to the telephony infrastructure.
Integration with Practice Management Systems
The AI agent is only useful if it can actually book appointments. We built a scheduling integration layer so the first version of Loquent could check real-time availability and create bookings directly in the clinic's calendar.
What We Shipped
By week 8, Loquent could:
- Answer inbound calls with a natural greeting customized per clinic
- Understand patient intent (booking, rescheduling, cancellation, general inquiry)
- Check real-time appointment availability
- Book, reschedule, and cancel appointments
- Collect new patient intake information
- Transfer to a human when the AI couldn't handle the request
- Handle calls 24/7, including evenings and weekends
Lessons Learned
Start with the hardest problem first. We tackled the voice latency challenge in week 1, not week 6. If we couldn't solve that, nothing else mattered.
Real users break everything. Our first live calls revealed edge cases we never anticipated — patients calling from noisy cars, thick accents, questions about parking. Each one taught us something.
Ship fast, iterate faster. We deployed to production at week 6 with a subset of features and spent weeks 7-8 iterating based on real call data. The live feedback loop was 10x more valuable than any internal testing.
What Happened Next
In August 2026 we rebuilt Loquent as an AI workforce for small businesses in Canada and the US. Owners now hire AI employees by role — AI Sales Coordinator, AI Customer Care Coordinator, AI Content Strategist and more — and the same employee works phone, website chat, WhatsApp, texts and email. Answering the phone became one job of many, and the voice pipeline described here became its phone channel. Loquent is now taking pilot businesses.
If you're building a voice AI product or want to bring AI voice agents to your healthcare practice, reach out to our team. We've done it before — in 8 weeks.