Build a Custom Voice AI for Candidate Prequalification

Recruiting teams often try generic AI interview platforms first. These tools are designed for behavioral questions like “Tell me about a time you faced a challenge,” but they cannot verify hard skills. They use broad keyword matching that fails when asked to confirm a specific, technical detail like a “TWIC card” or “6G pipe welding certification.” The AI can't distinguish a real certification from a candidate just mentioning the term.

Next, teams try building a phone tree with an IVR service. This approach is too rigid. An IVR follows a strict script and cannot handle natural conversation. If a candidate says, “My D1.1 is active but my D1.5 is pending,” the system marks it as a failure because it expected a simple “yes.” Candidate drop-off rates for these systems often exceed 70% because the experience is frustrating.

Connecting a web form to a Twilio voice message seems clever, but it’s brittle and expensive. The transcription task for a 5-minute call can cost over a dollar, and the logic can't handle dropped calls or retries. You end up with a high API bill for a system that provides incomplete transcripts and loses qualified candidates.

Syntora would begin an engagement by working with your team to map out the 5-7 most critical, non-negotiable pre-qualification questions. For example, in regulated industries like medical devices, this would often include specific FDA compliance knowledge or experience with particular submission types. This mapping would define the conversational flow and scoring rubric, which Syntora would implement using a Supabase table. This architecture allows your team to update questions and scoring logic directly without requiring new code deployments.

The core logic for the voice AI system would be built as a Python service using FastAPI. This service would manage the state of the conversation and interact with the Claude API for sophisticated natural language understanding. Syntora has extensive experience building document processing pipelines using Claude API for complex financial documents, and the same architectural pattern applies to interpreting nuanced candidate responses in this context. The system would be designed to interpret intent, not just keywords. For instance, a candidate stating “I’ve been running those submissions for five years” would be correctly scored as meeting a “3+ years of experience” requirement.

The FastAPI service would be deployed on AWS Lambda, an architecture chosen for its cost efficiency as you only pay for compute time during active calls. A webhook from your existing ATS (e.g., Bullhorn or Lever) would trigger the Lambda function when a candidate applies. Upon call completion, the system would post the full transcript, a clear pass/fail score, and a concise summary back to the candidate's record in your ATS. Syntora would implement robust error handling using httpx for external API calls, including automatic retries for transient network issues.

Syntora would configure structured logging using structlog, sending operational data to a central dashboard. This allows your team to monitor call volume, completion rates, and scoring accuracy. Customizable alerts, such as Slack notifications, could be configured to notify your team of significant changes, like a sudden drop in pass rates for a specific job requisition. The typical client input required would be specific domain knowledge for question design and access to relevant ATS APIs. Deliverables would include the deployed system, comprehensive documentation, and training for your internal team on its operation and maintenance.