Build a Custom Voice AI for Pre-Screening Candidates

Recruiting teams often try off-the-shelf video interview platforms like Spark Hire or MyInterview. These tools are expensive, charging per user or per job posting, and their scoring is generic. The AI gives a vague "strong communicator" score but cannot tell you if a candidate mentioned specific technical skills or followed a STAR-method response format, forcing you to watch the video anyway.

A more technical team might try to connect an audio transcription service like AssemblyAI to an LLM API using a low-code platform. This fails at the logic step. The workflow needs to transcribe audio, send the text to an API with a complex rubric, parse the JSON response, handle API errors with retries, and write structured data back to an ATS. This requires multiple conditional paths and error-handling branches that are brittle and expensive to run on a per-task basis.

For example, a logistics firm screening 150 dispatchers needs to check for experience with specific routing software. A generic tool can't do this. A low-code workflow attempting this becomes a tangled mess of duplicate steps, timing out on large audio files and offering no way to debug a bad score from the LLM.

Syntora would approach building a voice AI pre-screening system as a focused engineering engagement, starting with a discovery phase to define your specific requirements. The first step involves collaborating to define a detailed scoring rubric based on your essential screening questions. We would translate your ideal candidate answer criteria into a structured prompt for the Claude API, specifying the exact skills, keywords, and response structures for evaluation. This rubric would be designed as a configurable JSON file, allowing your team to update scoring parameters without developer intervention.

The core of the system would be a Python service, likely using FastAPI, exposing an endpoint to receive candidate audio files. This service would send the audio to AWS Transcribe, which provides text transcripts typically with less than a 10% word error rate. The transcript and the defined scoring rubric would then be passed to the Claude API. The Claude API is capable of returning a structured JSON object containing a score for each criterion, alongside a short rationale, usually within a few seconds. We have experience building similar document processing pipelines using Claude API for financial documents, and the same pattern applies to evaluating voice transcripts.

The entire system would be deployed on AWS Lambda, an environment well-suited for cost-effective, event-driven workloads. Candidates could record their answers through a simple web interface or via a Twilio-provisioned number. The engineered system would write the score, rationale, and full transcript to a custom note in your ATS. Syntora would develop custom integrations for specific ATS platforms such as Greenhouse and Lever, ensuring data flows correctly into your existing workflows.

Typical AWS and Claude API costs for processing around 500 candidates per month are projected to be under $50. The system would incorporate `structlog` for structured, queryable logs and configure CloudWatch alarms to provide alerts, for example, sending a Slack notification if API error rates exceed 1% over a five-minute period, to ensure operational stability. A typical build timeline for this level of complexity is 4-6 weeks, with an additional 2 weeks for integration and testing. Your team would need to provide access to your ATS API documentation and internal experts for rubric definition. Deliverables would include the deployed system, source code, and comprehensive documentation.