Build a Voice AI Recruiter for Your Niche

Recruiting teams often start with text-based screening questions in their Applicant Tracking System (ATS). This fails because many candidates write "see resume," and keyword matching cannot evaluate the nuance of real-world experience. You can't ask a follow-up question to a text box.

Traditional Interactive Voice Response (IVR) systems are the next logical step, but they are too rigid. Their "press 1 for yes" logic cannot handle open-ended questions. If a candidate says, "I have about five years of experience, but two of those were as a team lead," an IVR system either fails or incorrectly parses the response, making the data useless.

General AI call center platforms are designed for support or sales call summaries, not structured candidate evaluation. They might transcribe a call, but they lack the internal logic to score answers against a rubric, ask dynamic follow-up questions based on a previous response, and push a final, structured score into an ATS. Recruiters end up listening to every call recording, which defeats the purpose of automation.

Syntora's engagement would begin with a discovery phase to deeply understand your specific hiring needs. We would audit your existing top 3-5 job descriptions and current screening questions, collaborating to define the non-obvious, niche-specific skills essential for your target roles. Using the Claude API, we would analyze these materials to generate a precise set of 10-15 probing questions and a corresponding detailed scoring rubric. This initial configuration, defined through iterative refinement with your team, would be central to the system's effectiveness.

The core of the system would be a Python application, potentially deployed on AWS Lambda for scalability, designed to manage the conversational flow. When a candidate initiates a call via a dedicated phone number provisioned through Twilio, the application would orchestrate the interaction. It would use a high-quality text-to-speech engine to deliver questions and then transcribe candidate answers. The transcribed text would be forwarded to a FastAPI endpoint, which houses the primary scoring logic.

This scoring logic would involve a carefully structured prompt sent to the Claude API, instructing it to evaluate each answer against the established rubric on a defined scale (e.g., 1-5) and to identify key domain-specific concepts mentioned by the candidate. Based on this evaluation, the system would dynamically select the most appropriate next question from the pre-defined set, adapting the conversation in real-time. This dynamic questioning ensures a thorough assessment.

Upon completion of the call, a backend process would aggregate the individual question scores, generate a concise summary of the candidate's performance, and compile the full transcript of the conversation. The delivered system would then make an API call to your Applicant Tracking System (e.g., Greenhouse, Lever) to directly attach this comprehensive data as a note to the candidate's profile.

Building a system of this complexity typically involves a development timeline of 8-12 weeks. Key client deliverables would include the fully deployed and tested voice AI pre-qualification system, comprehensive documentation, and a transfer of ownership of the codebase. Clients would need to provide access to subject matter experts during the discovery and prompt engineering phases, and access to relevant APIs (like ATS integration points).