Stop Manually Reviewing Screening Calls. Use Voice AI.

Recruiting platforms like Humanly.io or Paradox offer AI screening, but they are built for large enterprises. For a small firm, their five-figure annual contracts and per-seat pricing are not feasible. Their workflows are also rigid; you cannot easily customize the summary format or the specific data points extracted from a call.

Many teams then try to build a partial solution using a transcription API like AssemblyAI or Deepgram. This approach fails because transcription is only 10% of the work. You get a wall of text, but a recruiter still has to spend 10-15 minutes reading it to find the candidate's salary requirements, notice period, and specific technical proficiencies. The core bottleneck, manual data extraction and entry, remains unsolved.

A 7-person staffing agency tried this with AssemblyAI. They piped call transcripts into their ATS, Bullhorn, but their recruiters still spent hours manually reading each one to fill out custom fields. They paid for an API but saw no meaningful reduction in time-to-submit, because the high-value summarization work was still completely manual.

Syntora's approach would begin with an audit of your existing call recording infrastructure. Call recordings, typically MP3 or WAV files from a VoIP system like Aircall, would be ingested into an AWS S3 bucket. An S3 event trigger would then invoke an AWS Lambda function. This function would send the audio to AWS Transcribe, configured for speaker diarization to separate the recruiter's and candidate's speech for cleaner analysis.

Once the raw transcript is obtained, the core logic would run in Python using the Claude API. Syntora would collaborate with your team to design a detailed prompt that instructs the model to act as a senior technical recruiter, extracting specific data points into a structured JSON format. This would include fields identified during discovery, such as 'salary_expectations_usd', 'key_skills_mentioned', 'relocation_status', and a concise summary of the conversation. This process is designed for efficient transformation of unstructured transcripts into clean, usable data.

The resulting structured JSON object would then be posted to your ATS. Syntora would develop a custom integration using Python's httpx library to interact with the REST API of your specific platform, whether it is Bullhorn, Crelate, or Greenhouse. The data would populate the candidate record automatically, typically attaching the summary as a note. The architecture is designed to support rapid processing from audio ingestion to ATS update.

For operational visibility, the FastAPI service that orchestrates these components would use structlog for structured JSON logs, which would be sent to AWS CloudWatch. Syntora would configure alerts that fire on specific error codes or processing issues, sending these alerts to a dedicated Slack channel for immediate investigation. The system would also incorporate robust error handling, including exponential backoff retry mechanisms for transient API errors.