An AI that interviews candidates — live, by voice, and scores them.
An HR technology company running screens for enterprise hiring teams came to us with a backlog: thousands of applicants a month, a dozen engineers doing first-round calls, and no two interviews alike. We built them an interviewer that reads the résumé, asks adaptive questions out loud, and hands back a scored summary within minutes of the candidate hanging up.
- Client
- HR technology company, enterprise hiring
- Users
- Recruiters, hiring managers, candidates
- Engagement
- Design, build, deploy, ongoing run
- Status
- Live in production
The Problem
“We were paying senior engineers to ask the same five questions all day.”
That was the opening line of our first call with the client. Their clients posted roles across engineering, finance, sales and support. Every role meant a wave of 30-minute first-round screens, and every screen pulled an engineer or a hiring manager off real work. Candidates in other time zones waited days for a slot. Some simply took another offer.
The quality problem was worse than the volume problem. Two interviewers, same candidate, different verdicts. Notes lived in personal docs or nowhere. When a hiring committee disagreed, there was no transcript to go back to, only memory.
They had tried a question-bank tool. Candidates recognised the questions from Glassdoor within a week. What they actually needed was an interviewer that reads this résumé for this job, runs the same way at 2am as at 2pm, and leaves a record a committee can trust.
Our Approach
Built around how their recruiters already worked.
Questions grounded in the résumé and the job
The candidate uploads a résumé; the recruiter pastes the job description. The platform extracts the required skills, then runs a hybrid flow: roughly 60% of questions follow what the candidate brings up, 40% are strategic picks that make sure every required skill gets covered. Round-robin topic selection stops it repeating itself.
A real conversation, not a form
Audio streams over a WebSocket in chunks to a self-hosted Whisper service for speech-to-text. A 3D avatar speaks the questions with lip-sync. Heartbeats every 15 seconds and exponential-backoff reconnects keep a 30-minute interview alive through flaky Wi-Fi. Mid-session token expiry is handled without dropping the call.
Cost control designed in, not bolted on
Every résumé and job description is hashed (SHA-256) and its summary cached in PostgreSQL, so the same documents never hit the AI twice — up to 80% fewer model calls on repeat screens. OpenAI is primary, Gemini is the automatic fallback, and interviews are billed per minute against pre-checked credits.
Inside the Product
What their recruiters and candidates see.
What Shipped
What they got, and what it cost them to run.
“What stands out is how naturally it conducts interviews — engaging, interactive, and remarkably close to the real thing. It’s much more than a tool that asks questions: it evaluates your knowledge, tests how you approach situations, improves how you articulate your thoughts, and shows you where to work on yourself. A brilliant use of AI to solve a real-world problem. Highly recommended.”
Have a project? Let's scope it in one call.
We reply within one business day.