Build a Custom Voice AI for Inbound Call Handling

Small service companies often start with a basic Interactive Voice Response (IVR) system from providers like RingCentral or 8x8. These are simple menus that force callers to navigate by pressing keys. They cannot understand a customer who says, "My basement is flooding." This leads to frustrated callers who hang up and call a competitor, especially in an emergency.

A more technical approach is using a tool like Twilio Studio. While powerful, its visual builder becomes unmanageable for conversations with more than a few turns. Logic like checking a Supabase table for customer status requires writing separate serverless functions, splitting your business logic between a visual diagram and scattered code. This setup is brittle and difficult to troubleshoot when a call flow fails.

Imagine a 15-person HVAC company using a standard IVR. A caller with a broken air conditioner on a 95-degree day must listen to a 45-second menu about business hours and promotions. All they want is to book a technician. The system cannot understand their urgent request, logs it as just another call, and the potential customer moves on. The per-minute fees are charged whether the call results in a booked job or a hang-up.

Syntora's approach to building a voice AI system for inbound call handling would begin with a discovery phase. We would work with your team to map out the 3 to 5 most common reasons people call your business, such as booking a new job, checking an appointment status, or asking a billing question. This step is crucial for defining the initial scope and intent recognition requirements.

Technically, the system would be architected around a phone number provisioned through Twilio, pointing to a webhook. This webhook would be a Python application built with FastAPI, designed to receive events for every incoming call. When a call is answered, the audio stream would be transcribed in near real-time. We have experience building document processing pipelines using Claude API for financial documents, and the same pattern applies here for handling conversational audio. The transcribed text would be sent to the Claude API to perform intent recognition and entity extraction, identifying what the caller wants and pulling out key details like their name, address, and the specific service they need.

Once the caller's intent is identified, the FastAPI application would execute the necessary action. For instance, if the intent is to book a new appointment, the system would make an API call to your existing scheduling software to find and book an available slot. A text-to-speech engine would then be used to confirm the details with the caller.

The delivered system would typically be deployed on serverless infrastructure such as AWS Lambda, which helps manage operational costs efficiently. To ensure observability, Syntora would integrate structured logging using a tool like structlog, sending all call data to a Supabase database. This would provide a searchable dashboard with a full transcript and outcome for every interaction, allowing you to monitor how the system performs. This approach ensures you own the full source code and can easily adapt the system as your business needs evolve.