Build a Custom Voice AI Recruiting System

Recruiting teams often try off-the-shelf applicant tracking systems (ATS) like Greenhouse or Lever that have 'AI' features. These tools are good for keyword matching on resumes but cannot conduct or evaluate a live phone conversation. They identify candidates who used the right words, not those who can articulate their experience.

Consider a 15-person staffing agency screening for a warehouse associate role. They use a tool like MyInterview, which provides pre-recorded video questions. The problem is that candidates can re-record answers, and recruiters still have to manually watch every 3-minute video. With 100 applicants, that is 5 hours of video review, which is no faster than doing phone screens.

Another approach is using API-driven services like AssemblyAI for transcription and then feeding the text to a GPT model. This breaks down because the raw transcript lacks context. The system cannot distinguish between a candidate's confident explanation and a hesitant, rambling answer. Without a fine-tuned model for conversational analysis, the scores are meaningless and recruiters do not trust the output.

Syntora would begin by analyzing your 3-5 key screening questions for a high-volume role. We would configure the Claude API to generate a structured interview script that includes follow-up probes for incomplete answers. Candidate audio files and their transcriptions would be stored in a Supabase database, providing a single source for data and potential retraining.

The core system architecture typically involves a FastAPI application deployed on AWS Lambda. When a candidate calls the dedicated number, the service would record their responses. An audio processing library would clean the recording before sending it to a transcription service. The full transcript, which we estimate would generate in about 90 seconds for a 5-minute call, would then be passed to a scoring module.

The scoring module would be a Python script that uses the Claude API with a detailed rubric prompt. This module would score each answer from 1-5 on criteria like clarity, relevance, and experience. The final report, which Syntora would configure for delivery as a PDF and a webhook to your ATS, would include the full transcript, individual scores, and a summary score. The target processing time from call end to report delivery would be under 3 minutes, depending on call length and API latency.

A system built with this architecture could operate for under $40 per month in usage costs for up to 500 screenings. By using serverless functions via AWS Lambda, the architecture ensures that compute time is only paid for when a screening is actively processed. This approach avoids the higher monthly fees of SaaS platforms that often charge per seat, regardless of usage.