Build a Custom Voice AI for Automated Reference Checks

Recruiting teams often start with email-based reference tools like Checkster. These are essentially automated surveys. The problem is that references provide brief, low-effort written answers, and response rates are often below 50%. There is no way to ask a follow-up question or gauge the tone and conviction behind an answer, which is where the real signal is.

A more significant issue is the manual process itself. A recruiter at a small firm trying to fill 15 roles a quarter must conduct over 100 individual reference calls. Each 20-minute call is preceded by 10 minutes of email tag for scheduling. This consumes more than 50 hours per quarter, time that could be spent sourcing new candidates. The process is impossible to scale without hiring more recruiters.

Some firms try to solve this with virtual assistants, but that introduces new problems of consistency and quality control. One VA may be a great interviewer while another just reads the script. This introduces significant bias and noise into the hiring process. The data collected is unstructured and difficult to compare across candidates, making the entire effort subjective.

Syntora would begin an engagement with a discovery phase, auditing your existing Applicant Tracking System (ATS) and specific hiring workflows. This would define the integration points and requirements for question sets, which would be managed and stored in a Supabase database, allowing for tailored questions per job role.

The core of the system would be a Python application built with FastAPI. This application would utilize Twilio to place outbound calls to references. For real-time speech-to-text transcription, Syntora would integrate Deepgram, chosen for its high accuracy potential on phone lines, enabling the conversational AI. The conversational flow would be managed by the Claude 3 Sonnet API, processing transcripts and dynamically formulating follow-up questions to maintain a natural dialogue.

Upon call completion, the full transcript would be routed to the Claude 3 Opus API for detailed analysis. This model would extract key information, score the reference's feedback against predefined competencies from the job description, and generate a concise summary. This analysis would produce a structured JSON object and a PDF summary report. Syntora has experience building document processing pipelines using Claude API for sensitive financial documents, and the same pattern applies to synthesizing feedback from reference calls.

The entire service would be designed for efficient deployment, often using serverless architectures like AWS Lambda. This approach ensures the system scales with demand and optimizes operational costs. The final PDF report and structured data would be pushed back into the candidate's profile within your ATS via its API, and notifications could be configured for recruiting teams, such as via Slack, to alert them when results are available. A typical build of this complexity could be delivered within 8-12 weeks, depending on the complexity of ATS integration and analysis requirements. The client would provide access to their ATS APIs, define question sets, and participate in iterative feedback sessions.