Build a Custom Voice AI for Automated Reference Checks

Recruiting teams often start with manual phone calls for reference checks. This process is slow and inconsistent. One recruiter might ask probing follow-up questions while another sticks to the script, making it impossible to compare candidates objectively. A single senior candidate requiring five reference calls can consume half a day of a recruiter's time.

Trying to scale this, teams turn to survey tools like Typeform or built-in ATS features. These just send an email with a link to a form. Response rates are often below 30% because it puts the work on the reference. A reference who would happily talk for 10 minutes will ignore a form that looks like a 20-minute task. There is no way to ask dynamic follow-up questions based on an initial answer.

Dedicated reference checking platforms like Checkster exist, but they charge per-candidate or per-seat, which is expensive for a small business. A 5-person recruiting team can face a bill of over $500/month. These platforms also offer limited customization of the analysis logic, flagging generic risks instead of the specific competencies you screen for.

Syntora's approach to building a voice AI reference checking system begins with a discovery phase to codify your existing reference questions and decision criteria into a structured script. We would use Anthropic's Claude 3 Sonnet API to manage the conversational flow, enabling dynamic follow-up questions that delve deeper into a reference's initial response. This conversational logic would be deployed as a Python service using FastAPI.

When a recruiter triggers a check from your ATS, a webhook would fire an AWS Lambda function. This function would use the Twilio API to place the outbound call. Audio streams in real-time to AWS Transcribe for speech-to-text conversion. The transcribed text is then sent to the Claude API, which generates the next question. The response is converted back to audio using AWS Polly and played to the reference. This question-response loop typically completes within a second.

All call data, including full audio recordings, transcripts, and a final structured summary, would be written to a Supabase database. The summary, which can include flags for hesitation, negative sentiment, or conflicting information, would be posted back to a custom field in your ATS via its API. Recruiters would receive a complete report shortly after the final call ends. Typical cloud services cost for hundreds of calls per month would be modest, often under $50.

As part of the engagement, Syntora would build a simple dashboard showing call completion rates, average call duration, and common keywords from transcripts. The system would include structured logging with structlog, configured to send alerts to Slack via webhooks if a call fails due to a technical issue or if transcript quality scores are low, indicating a bad connection. This allows for immediate review.

A typical engagement to design and build a system of this complexity ranges from 8 to 12 weeks. Clients would need to provide access to their ATS documentation, existing reference questions, and participate in regular feedback sessions. Deliverables would include the deployed system on client infrastructure, source code, detailed architectural documentation, and a user guide.