Build a Custom Voice AI for Applicant Questions

Most small businesses start with a standard phone tree (IVR) from their VoIP provider. These systems fail because they are not conversational. An applicant asking, "Did you get my application for the forklift operator job?" cannot be answered by a "Press 1 for hours" menu. Callers inevitably press '0', defeating the purpose of the system.

A dedicated contact center platform like Talkdesk offers AI features, but they are designed for 100-plus agent teams and carry a per-seat monthly cost that is unfeasible for a small recruiting team. The setup and maintenance also require a dedicated IT specialist, which most small businesses lack. The alternative, trying to connect a web-based chatbot like Drift to a phone line, results in a frustrating experience with 2-3 second delays between speaking and hearing a response, making a natural conversation impossible.

This leaves the office manager or a junior recruiter spending hours each day answering the same repetitive questions: "Are you still hiring?", "What's the status of my application?", "What are the benefits?". This manual effort prevents them from focusing on sourcing and interviewing qualified candidates. The result is significant: a team spending 10 hours per week on repetitive phone calls can typically reduce that to under two hours with a well-built voice agent.

Syntora's approach to integrating voice AI for inbound applicant queries begins with a discovery phase. We would identify the 20-30 most frequent questions your applicants ask and map these to relevant data fields within your ATS, such as Lever or Greenhouse. This process informs the creation of a structured knowledge base, which would be stored in a Supabase Postgres database, allowing the AI to query for static information like office locations or benefits summaries.

For telephony, the system would utilize Twilio to manage the phone number and handle the incoming call stream. When a person calls, their speech is converted to text via Twilio's capabilities and sent to a custom FastAPI application. This application would then send the transcribed query and conversation history to the Claude API, which generates a natural language response. The entire process is designed for low latency, aiming for a swift conversational turn.

The AI's text response would be converted back to audio using AWS Polly, a text-to-speech service known for its low latency. This audio would then be streamed back to the caller through Twilio. For dynamic queries that require real-time data, such as application status checks, the AI agent would employ a tool-calling function. This function would invoke an AWS Lambda function to securely query your ATS API. We have experience building robust API integrations and ensuring data security in sensitive contexts.

The core FastAPI service would be deployed as a Docker container on AWS App Runner, providing a managed and highly available environment. The system would log every call transcript and AI response to your Supabase instance, facilitating review and continuous tuning. We would implement structured logging with `structlog` to monitor performance, configuring alerts to a shared Slack channel if API latency from the ATS exceeds defined thresholds or error rates spike. Typical hosting costs for this architecture are generally under $50/month. A typical build timeline for a system of this complexity, from discovery to initial deployment, would be approximately 6-10 weeks, requiring client involvement for ATS access, knowledge base content, and testing.