Build a Custom System to Prequalify Recruiting Candidates

Most recruiting teams first try off-the-shelf AI interview platforms. These tools seem attractive because they offer a quick setup, but their business models fail small businesses. They typically charge per seat or per interview, meaning a 10-person firm could pay over $1,500 per month for a single function, with costs ballooning during a hiring surge.

A more critical failure is their rigid logic. These platforms let you customize questions but cannot dynamically branch the conversation. For example, a technical recruiting firm needs to screen software engineers. Their platform asks, "Do you have experience with Python?" It can score a "yes" answer, but it cannot ask a follow-up like, "Describe a time you used an async library like httpx." This means a junior developer who took one online course gets the same score as a senior engineer, creating hours of wasted effort for human recruiters who must re-qualify everyone.

These platforms are designed for high-volume, generic screens, not for the nuanced qualification that small, specialized firms depend on. They cannot distinguish shallow keyword matches from deep expertise, which is the entire point of a technical screen. The result is a noisy, unreliable signal that damages recruiter productivity.

Syntora's approach to Voice AI pre-qualification begins with a discovery phase to map your existing screening questions into a decision tree and identify primary applicant sources, whether an ATS like Lever or dedicated inboxes. The system would use the Claude API for its advanced conversational abilities and precise instruction following to design an effective interview flow. The complete script, along with its branching logic, would be stored in a Supabase database, allowing for flexible question editing without direct code changes.

The proposed system's core would be a Python application, likely built with FastAPI. When a new candidate applies, a webhook would trigger an AWS Lambda function, which would initiate the automated call. A dedicated voice API would handle real-time transcription and speech synthesis. The system's logic would operate as a state machine, pulling the next question from Supabase based on the candidate's transcribed previous answer to enable dynamic conversations.

After a call completes, typically designed to run for 90 to 120 seconds, the full transcript would be sent back to the Claude API. A carefully engineered prompt would extract key information such as years of experience, salary expectations, specific skill mentions, and potential red flags, outputting this into a clean JSON object. This structured data would then populate custom fields in your ATS or a new row in a designated spreadsheet. The post-call processing pipeline would be engineered for rapid execution, with a target completion time under 8 seconds.

Syntora would deliver the complete, commented source code to your private GitHub repository and assist with deploying the system on your chosen cloud infrastructure. We would configure structured logging using structlog and establish alerts that notify administrators via Slack if API error rates exceed a defined threshold or if transcription accuracy falters. For a system processing up to 1,000 candidates monthly, estimated cloud hosting costs typically range under $75. We have experience building document processing pipelines using Claude API for financial documents, and the same architectural patterns apply effectively to documents relevant to recruitment.