Automate Recruiting Reference Checks with a Custom Voice AI

Recruiting teams start with manual phone calls, which are slow and inconsistent. Different recruiters ask questions differently, and hand-written notes are impossible to search or compare across candidates. This process creates a bottleneck, delaying offers and causing candidates to accept competing ones.

Off-the-shelf software like Checkster or Outmatch tries to solve this but relies on sending web forms to references via email. This approach suffers from low response rates as references ignore the emails. The written answers they do collect lack the nuance and candor of a spoken conversation. These platforms also charge per-hire or per-seat fees that become expensive for a growing firm.

Technically-inclined teams might try a visual IVR builder like Twilio Studio. While it can place calls and play audio prompts, it cannot handle the critical task of understanding unstructured verbal responses. A simple "press 1" workflow is easy, but analyzing a two-minute answer about a candidate's weakness requires a custom speech-to-text and language model pipeline, which cannot be built within a drag-and-drop interface.

Syntora would approach the development of your custom voice AI reference checking system through a phased engagement. We would begin with a discovery process to thoroughly map your precise reference questions into a structured script, preparing it for automated delivery. Simultaneously, we would provision a dedicated phone number using the Twilio API, integrating it directly into the proposed system architecture.

A critical component of the engagement would be establishing direct integration with your Applicant Tracking System. Whether your organization uses Greenhouse, Lever, or Ashby, we would leverage its API and develop a robust Python script with httpx to automatically pull candidate and reference contact details. This foundational setup is designed to eliminate manual data entry for your recruiting team, streamlining the entire workflow.

The core of the system would be a custom Python application built with the FastAPI framework, engineered for scalability and reliability. For event-driven execution and optimal resource utilization, it would be deployed on AWS Lambda. When a candidate reaches the designated reference check stage in your ATS, a webhook would trigger the Lambda function. This function would then place outbound calls, play pre-recorded audio prompts for each question, and record the reference's spoken responses, allowing for configurable durations per question.

Each audio recording would be sent to Anthropic's Claude 3 Sonnet API for advanced natural language processing. Syntora has extensive experience building document processing pipelines using Claude API for sensitive financial documents, and the same robust pattern applies to analyzing spoken responses for reference checks. We would craft a specific prompt chain to instruct the model to perform multiple tasks: generating a verbatim transcript, writing a concise summary of the entire call, and identifying predefined keywords or phrases relevant to your hiring criteria, such as "conflict" or "leadership."

The resulting structured data, including the summary, full transcript, and a secure link to the original audio recording, would then be posted as a private note on the candidate's profile within your ATS, utilizing the ATS API. The system's architecture would be designed to efficiently handle concurrent calls and scale with your organization's volume. While specific monthly hosting costs would vary based on actual usage, components like AWS Lambda and Supabase are selected for their cost-effectiveness in processing substantial numbers of calls.

Our engagement would conclude with the delivery of the fully functional, tested application within your cloud environment, complete with the entire source code, detailed technical documentation, and hands-on knowledge transfer sessions for your internal teams. Your organization would be instrumental in providing necessary ATS API credentials, defining the specific reference questions, and validating system outputs throughout the development and testing lifecycle.