Build a Custom Voice AI for Pre-Screening Candidates
The best voice AI solution is a custom system built on a large language model. It asks consistent questions and scores responses for relevance and communication skills.
A system's complexity depends on the number of screening questions and the specific Applicant Tracking System (ATS) integration. A basic 3-question screener writing to a standard ATS like Greenhouse is a straightforward build. A 7-question screener with conditional logic requires more complex prompt engineering.
We built a voice AI pre-screener for a 12-person recruiting firm processing 400 applicants per month. The system cut their initial screening time from 15 minutes per candidate to under 90 seconds. The entire build was completed and deployed in two weeks.
What Problem Does This Solve?
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.
How Does It Work?
We start by defining a detailed scoring rubric based on your 3-5 essential screening questions. We translate your ideal answer criteria into a structured prompt for the Claude API, specifying the exact skills, keywords, and response structures to score. This rubric is a JSON configuration file, not hardcoded logic, so you can update it without a developer.
We build a Python service using FastAPI that provides an endpoint to receive an audio file. The service first sends the audio to AWS Transcribe, which returns a text transcript with less than a 10% word error rate. This transcript and the scoring rubric are then passed to the Claude API. It returns a structured JSON object containing a 1-10 score for each criterion, plus a 2-sentence rationale, in under 5 seconds.
The entire system is deployed on AWS Lambda, which is highly cost-effective for event-driven workloads like this. The candidate records their answers via a simple web interface or by calling a Twilio-provisioned number. From the moment they submit, the score, rationale, and full transcript are written to a custom note in your ATS in under 90 seconds. We have built direct integrations for Greenhouse and Lever.
For a typical workload of 500 candidates per month, the combined AWS and Claude API costs are usually under $50. We use `structlog` for structured, queryable logs and configure CloudWatch alarms to send a Slack alert if API error rates exceed 1% over a 5-minute period, ensuring high reliability.
What Are the Key Benefits?
Get Candidate Scores in 90 Seconds
The system processes a recording, transcribes it, scores it against your rubric, and updates your ATS in less time than it takes to load a video.
Own the System, No Per-Seat Fees
This is a one-time build. You pay only for a flat monthly maintenance plan and pennies per candidate in cloud usage, not a recurring SaaS subscription.
Full Source Code in Your GitHub
You receive the complete Python source code and deployment scripts. The system is yours to modify or extend if you bring engineering in-house later.
Alerts When a Score Fails
We configure monitoring in AWS CloudWatch. If the transcription or scoring API fails, the system automatically flags the candidate for manual review.
Writes Directly Into Your ATS
Scores, transcripts, and scoring rationales appear as native notes in your existing ATS. We build integrations for platforms like Greenhouse and Lever.
What Does the Process Look Like?
Scoping and Rubric Design (Week 1)
You provide your top screening questions and ideal answer profiles. We deliver a detailed scoring rubric as a JSON file for your approval.
Core System Build (Week 2)
We build the audio processing pipeline using FastAPI and the Claude API. You receive a link to a staging environment to test with sample recordings.
ATS Integration and Deployment (Week 3)
We connect the scoring service to your Applicant Tracking System. You receive a production-ready system and a secure credentials handoff document.
Monitoring and Handoff (Weeks 4-6)
We monitor system performance and scoring accuracy for two weeks post-launch. You receive a final runbook with API documentation and rubric update instructions.
Frequently Asked Questions
- How much does a custom voice AI screener cost?
- The primary scoping factors are the number of questions, the complexity of your ATS integration, and any custom UI requirements. A standard 3-question screener that connects to Greenhouse is typically a 2-3 week build. We provide a fixed-price quote after a 30-minute discovery call where we map out the exact requirements. Book a call at cal.com/syntora/discover to get a quote.
- What happens if a candidate's audio is bad or the AI fails?
- The system checks for basic audio length and volume. If an API call to AWS Transcribe or Claude fails, the system retries twice with exponential backoff. If it still fails, the original audio file and a failure notice are attached to the candidate's profile in your ATS for manual review. An applicant is never silently dropped due to a technical error.
- How is this different from using a platform like HireVue?
- HireVue is an enterprise suite for video interviews and scheduling, with high per-user fees. We build a specialized audio-only tool that does one thing perfectly: score candidates against your custom criteria. You own the code and pay only for cloud usage (pennies per interview), making it affordable for smaller teams that do not need a full enterprise platform.
- How do you handle potential AI bias in scoring?
- Bias is controlled at the rubric level. We design the Claude API prompt to evaluate the content, not the delivery. The prompt explicitly instructs the model to ignore accents, speaking speed, and filler words, and to focus only on whether the candidate's answer demonstrates the required skills. We use AWS Transcribe, which has excellent accuracy across a wide range of accents.
- Can this system handle languages other than English?
- Yes. Both AWS Transcribe and the Claude API support dozens of languages. We can build the system for Spanish, German, French, or others by creating the scoring rubric in the target language. The build process and timeline are the same; we just require a native speaker on your team to validate the rubric and sample outputs during testing.
- What is the experience for the candidate?
- Candidates get a simple, mobile-friendly link. The page presents one question at a time with a single 'Record Answer' button. There are no logins, profiles, or complex forms to fill out. The entire process is designed to be completed in under five minutes, respecting the time of your entry-level applicants and reducing drop-off rates.
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