Build a Custom Voice AI Recruiting Screener for Service Firms

Most recruiting teams try SaaS tools like MyInterview or Willo for initial screening. These platforms fail because their scoring logic relies on simple keyword matching. A candidate who mentions 'Python' gets points, but the system cannot distinguish between ten years of production experience and a three-month bootcamp. The scoring is too shallow to be reliable.

This forces recruiters to listen to every single audio recording anyway, which defeats the purpose of the tool. They end up paying a $499/month subscription for a system that creates more work. The core issue is that these tools cannot conduct a dynamic conversation or understand the nuance in a candidate's answer. They are a collection of recorded questions, not an intelligent interviewer.

Larger ATS platforms like Greenhouse offer AI add-ons, but these are bundled into enterprise tiers costing upwards of $25,000 annually. For a small or medium-sized business, this is not a viable option. You are forced to choose between a cheap tool that does not work or an expensive platform with features you will never use.

Syntora's engagement would begin by collaborating with your team to define and analyze your most critical screening questions for a specific role. We would then use the Claude API to refine these questions and develop a strategy for generating logical follow-up prompts based on candidate responses. The core system architecture would be a Python application built with FastAPI, designed for deployment on AWS Lambda to ensure efficient, consumption-based execution that scales with your needs. Syntora has experience building similar document processing pipelines using the Claude API for financial documents, and the same architectural patterns apply to voice transcript analysis.

When a candidate calls the dedicated Twilio number, a webhook would trigger the Lambda function. The system would record their audio, transcribe it using a fine-tuned speech-to-text model, and then send the transcript to the Claude API. The prompt would be engineered using your specific role requirements and a detailed scoring rubric developed jointly. Claude would then score the answers across key criteria we define together, such as technical depth or communication clarity, providing a written justification for each score.

The final output would be a JSON object containing the full transcript, granular scores, and a summary of the candidate's fitness. This object would be posted directly to your Applicant Tracking System as a note on the candidate's profile. We would use httpx to build direct API integrations for platforms like Greenhouse, Lever, or Ashby, ensuring data flows into your existing tools.

Supabase would be used to store a log of every screening for accuracy analysis and future prompt tuning. All logging would be handled by structlog for machine-readable output, with alerts configured to post to a dedicated Slack channel if transcription confidence drops below a defined threshold or the Claude API returns an error. This infrastructure typically costs under $50 per month for up to 1,000 candidate screenings. A typical build for this complexity would take 8-12 weeks. The client would need to provide defined screening questions, access to their ATS API documentation, and participate in regular feedback sessions. Deliverables would include the deployed voice screening system, detailed architectural documentation, and a training session for your team.