Automate Your Reference Checks with a Custom Voice AI

Recruiting teams often try using generic calling services. A tool like Twilio Studio can place calls, but it cannot interpret the unstructured audio responses or summarize the results into a candidate profile. You get a raw MP3 file, which a recruiter still has to listen to, defeating the purpose of automation.

A common scenario involves a 15-person firm trying to scale from 20 to 50 hires a year. Their two recruiters are a bottleneck. Reference checks for 5 finalists take 10-15 hours of phone tag and note-taking. They look at enterprise recruiting platforms like HireVue, but the $20,000 annual contract and per-seat licensing model are designed for 500-person companies, not them.

These off-the-shelf tools fail because the core problem is not placing the call; it is understanding the answer. They force a choice between a raw audio file that creates more manual work or an expensive, inflexible AI module that cannot be customized to ask role-specific questions. They provide a partial solution that does not address the actual bottleneck: recruiter time spent on low-value, repetitive conversations.

Syntora's approach to building a custom voice AI reference checking system would begin by mapping your existing reference questions into a structured script. We would leverage the Claude API to generate dynamic, natural-sounding follow-up questions based on the reference's initial answers. This ensures a more adaptive conversation than simple static scripts. We have experience building similar document processing pipelines using Claude API for financial documents, and the same pattern applies effectively to extracting insights from voice interactions.

The system would integrate with your existing ATS (e.g., Lever, Greenhouse) using Python and the httpx library to pull candidate and reference contact information automatically. The core calling logic would be built on AWS Lambda, utilizing Twilio's API for programmable voice. When a candidate reaches the reference check stage in your ATS, a webhook would trigger the Lambda function. This function would initiate an outbound call, ask the scripted questions, and record the reference's responses. The audio file, typically 2-4 minutes long, would be saved directly to a secure Amazon S3 bucket within your AWS account.

A second AWS Lambda function would be configured to trigger the moment the audio file is saved. It would use an audio-to-text model to produce a full transcript, a process that typically takes about 15 seconds for a 3-minute call. That transcript would then be sent to the Claude API with a carefully engineered prompt sequence. This prompt would instruct the AI to generate a concise summary, score responses against key competencies, and extract any predefined keywords relevant to your hiring criteria.

The entire analysis pipeline would be designed for efficient execution. The final summary, competency scores, and full transcript would then be posted back to the candidate's record in your ATS via an API call. We would implement robust logging using structlog, and design automated alerts (e.g., to a Slack channel) for failed calls or low-confidence transcriptions (below 90%). The architecture would be designed for scalability, supporting concurrent calls as needed, with typical AWS hosting costs for such a system estimated to be under $50 monthly, depending on volume.