Build a Voice AI System to Prequalify Candidates

Recruiting teams often start with off-the-shelf voice AI platforms. These tools are rigid, designed for simple yes/no questions, and cannot handle the nuanced, multi-part answers required for real screening. They can ask "Do you have 3 years of experience?" but fail when a candidate says, "I have two and a half years at my current job and one year at my previous one." The system either fails or just saves a useless transcript, forcing a manual review.

A common failure scenario involves a nursing recruitment agency trying to pre-qualify candidates. Their key question is "Tell me about your nursing licenses, including state, number, and expiration date." A SaaS tool can only transcribe the audio. This leaves the recruiter to listen to 200 voicemails a week, defeating the purpose of automation. On top of that, per-call pricing of $0.75 adds up to $300/month for a glorified answering machine.

Some teams attempt a DIY solution using Twilio and a basic speech-to-text API. They quickly find that a raw transcript is not useful data. The real challenge is Natural Language Understanding: parsing the transcript, extracting structured entities like dates and license numbers, and scoring the answers. This is where DIY projects stall, burning engineering hours without producing a usable system.

Syntora's engagement would begin by collaboratively mapping your 5-7 core pre-screening questions to a strict data schema, often leveraging Python's Pydantic library. This defines the precise JSON object that the system would generate and post to your Applicant Tracking System (ATS), ensuring consistent candidate summaries. A dedicated phone number for the agent would be provisioned using the Twilio API.

The proposed core architecture would leverage a FastAPI application deployed on AWS Lambda for event-driven execution. When a candidate calls, Lambda would invoke the application. For the conversational engine, Syntora would integrate the Claude API. This allows the system to understand the candidate's full response, ask clarifying questions if an answer is ambiguous, and maintain context throughout the call.

Once a call concludes, the system would make an asynchronous call to Claude using libraries like httpx to summarize the entire conversation and score it against your predefined rubric. This process generates a structured JSON payload containing the candidate's answers, a concise summary, and an overall qualification score. This payload would then be posted to your ATS's REST API. Transcripts and audio would be stored in a Supabase Postgres database for a defined retention period.

For operational visibility, Syntora would implement structured logging, potentially using structlog, with logs sent to AWS CloudWatch. Proactive monitoring would be configured to trigger alerts based on critical metrics like API error rates or significant increases in call processing times, ensuring system health and performance. We have built document processing pipelines using Claude API for financial documents, and the same pattern applies effectively to pre-screening recruitment documents and conversations.