Custom Voice AI for Screening Entry-Level Applicants
Off-the-shelf platforms like MyInterview or HireVue work well for standard, high-volume roles. Custom-built systems are better for roles requiring nuanced technical or cultural evaluation.
We built a voice screening system for a 12-person recruiting firm that processes 400 junior developer applicants per month. The system cut their initial screening time from 25 minutes per candidate to just 4 minutes and went live in a 3-week build cycle.
A pre-built platform gets you started in hours but uses generic scoring models that assess tone and confidence. A custom system is tuned to your specific job descriptions and success criteria, scoring candidates on the substance of their answers, not just their delivery.
What Problem Does This Solve?
Recruiting teams often start with automated interview platforms like Spark Hire or Willo. These tools are great for replacing one-way video screens but fail at accurately assessing technical roles. Their scoring models are black boxes, often grading for generic traits like 'clarity' which has little correlation with a developer's ability to write Python code.
A specialized recruiting firm tried using one of these platforms to screen for 10 different engineering roles. The platform could not differentiate a good answer about AWS services from a bad one. Recruiters had to listen to every 15-minute recording to manually score technical competence, which completely defeated the purpose of automation. They were paying $120 per recruiter per month for a glorified audio file host.
Furthermore, these platforms offer limited integrations. They might push a candidate's overall score to a major ATS like Greenhouse, but they cannot sync detailed, per-question feedback to a custom field or connect to an industry-specific recruiting system. This forces a workflow of manual copy-pasting that introduces errors and wastes time.
How Does It Work?
Our process starts by mapping your existing screening questions and scoring rubric directly into a set of prompts for the Claude API. We then build a simple API endpoint using Python and FastAPI that accepts an audio file from your careers page or application form. This endpoint is deployed as a serverless function on AWS Lambda to handle unpredictable applicant volume without ongoing server costs.
When an audio file is received, the Lambda function first sends it to Deepgram's transcription API. We use Deepgram for its high accuracy on technical jargon, achieving over 95% word accuracy on terms like 'PostgreSQL' and 'Kubernetes'. A typical 3-minute applicant response is transcribed into text in under 4 seconds.
Next, the raw transcript is sent to the Claude 3 Sonnet API. The prompt instructs the model to score the candidate's answers against the specific rubric we designed in step one, rating each response from 1 to 10 on criteria like technical correctness and problem-solving approach. The API returns a structured JSON object with scores, a summary, and direct quotes supporting the evaluation in about 9 seconds.
Finally, this JSON output is stored in a Supabase database for historical analysis, and a separate async function using httpx makes an API call to your specific ATS. It populates custom fields with the per-question scores and summary. The total processing time from audio submission to ATS update is under 15 seconds, and hosting costs are typically under $30 per month for 500 applicants.
What Are the Key Benefits?
Go Live in 3 Weeks, Not 3 Quarters
From kickoff to a production-ready system integrated with your ATS in 15 business days. Begin screening candidates automatically next month.
Pay Per Candidate, Not Per Recruiter Seat
A one-time build fee and low per-use API costs. Your expenses scale with applicant volume, not how many recruiters are on your team.
You Own the Code and the Scoring Model
The complete Python source code and all prompts are delivered to your company's GitHub repository. You have full control and no vendor lock-in.
Alerts When Transcription Quality Drops
We build monitoring that sends a Slack message if transcription confidence falls below 90%, flagging the recording for immediate human review.
Connects to Any ATS with an API
We write custom API connectors for your specific system, whether it's a major platform like Lever or an in-house recruiting database.
What Does the Process Look Like?
Week 1: Rubric and Architecture Design
You provide 3 target job descriptions and current screening questions. We deliver a finalized scoring rubric and a complete system architecture diagram.
Week 2: Core Pipeline Construction
We build the audio processing pipeline with FastAPI on AWS Lambda and integrate the transcription and analysis APIs. You receive a staging URL for testing.
Week 3: ATS Integration and Deployment
We write the custom connector to your Applicant Tracking System and deploy the full system to production. You receive API keys and documentation.
Weeks 4-8: Monitoring and Handoff
We monitor the first 200 live screenings, tune prompts for accuracy, and document maintenance procedures. You receive a final runbook.
Frequently Asked Questions
- How much does a custom voice screening system cost?
- The price is a fixed-fee build based on complexity. Key factors include the number of unique job roles to support, the complexity of your scoring rubric, and the quality of your ATS's API documentation. After the build, you only pay for API usage and minimal cloud hosting. Book a discovery call to discuss a specific quote.
- What happens if an applicant has a heavy accent or bad audio?
- The transcription API provides a confidence score for every word. If the average confidence for a recording falls below a set threshold (e.g., 85%), the system automatically flags it. It bypasses the AI analysis and creates a task in your ATS for a recruiter to review the audio manually, ensuring no candidate is unfairly penalized for poor recording quality.
- How is this different from a platform like MyInterview?
- MyInterview and similar platforms focus on personality and communication style using black-box models. Our approach builds a 'glass-box' system tailored to you. It scores candidates on specific, technical, role-based knowledge using a rubric you define and control. It evaluates what they say, not just how they say it, which is more predictive for technical roles.
- How is candidate data privacy handled?
- Candidate data is processed in-memory and never stored on Syntora systems. The audio files and results are stored in your own cloud infrastructure (your AWS account, your Supabase database). We can add logic to automatically redact Personally Identifiable Information (PII) from transcripts before they are sent for analysis or storage to meet compliance requirements.
- Can we change the screening questions ourselves after the build?
- Yes. The questions and scoring criteria are stored in a simple configuration file or a database table. Your team can edit these without needing to change any Python code. The runbook we provide includes instructions on how to update these configurations and safely redeploy the changes to the live system.
- How do you ensure the AI scoring is fair and unbiased?
- We start by building a 'golden set' of 20-30 screenings scored by your best recruiters. We then tune the AI's prompts until its scores consistently match your team's evaluations. After launch, the system periodically flags a random sample of interviews for human review, allowing you to audit the AI's performance and prevent model drift.
Related Solutions
Ready to Automate Your Small Business Operations?
Book a call to discuss how we can implement ai automation for your small business business.
Book a Call