AI Interview Platform & ATS
An AI-led hiring pipeline: résumé and job-description parsing, a real-time voice interview agent, and structured reports that let a recruiter review a candidate in minutes instead of sitting through every interview.
Python · real-time voice AI · LLMs · speech-to-text / text-to-speech
What the platform does
The platform takes hiring from a pile of applications to a short, structured review. It parses résumés and the job description, runs a live voice interview with each candidate, and produces a report. The recruiter makes the hiring decision; the platform's job is to make that decision quick and well-informed.
- IntakeRésumé and JD parsing
- InterviewLive AI voice interview
- TranscriptAccurate record of the session
- ReportStructured evaluation and integrity flags
- Recruiter decidesReviews the report and makes the call
Pipeline overview. The last step is always a person.
This is a proprietary IoTides product, so this page covers the problems, approach and my contribution, not internal architecture, vendors or figures.
My contribution
What I worked on
- Real-time interview agent: R&D and development of the live voice conversation, from the first prototype to the production version.
- Re-platforming: moved the session off a hand-built real-time prototype onto purpose-built real-time infrastructure.
- Conversation quality: tuning turn-taking, noise handling and recognition of domain terms.
- Latency instrumentation: stage-by-stage timing for every conversation turn.
- Session reliability: recording and failure handling for long interviews.
Résumé/JD parsing, the ATS and the wider product are team efforts at IoTides.
From prototype to production footing
The first version was a hand-built prototype. It proved an AI could hold a live interview, and showed that a home-grown real-time layer wouldn't hold up at production scale: media transport, reconnection, turn-taking and concurrency were all being reinvented.
A research pass into how production voice AI is built led to re-platforming the session onto purpose-built real-time infrastructure. After that, effort went into the conversation itself instead of keeping the plumbing alive.
Teaching the agent to listen like a person
An interview is only credible if the agent lets the candidate finish and hears the answer correctly. Most of the tuning went into that:
- Turn-taking: tuned so the agent stops cutting candidates off mid-sentence, without leaving an awkward pause afterwards.
- Real-world audio: keeping background noise and other voices out, because candidates join from wherever they are.
- Domain terms: priming recognition with the vocabulary each interview expects, so names and technical terms come through correctly.
- Exact transcripts: resolving streaming speech into one accurate record, so the report is built on what was actually said.
Making “it feels slow” measurable
“It feels slow” isn't a bug report. Every conversation turn is timed stage by stage, so the silence between a candidate finishing and the agent replying can be traced to a specific stage instead of guessed at.
That breakdown showed where the delay actually came from, which is what made it fixable: the changes it guided reduced the pause between turns. The same per-turn data also surfaces failures that are invisible from the outside, so they're caught in logs rather than in a candidate's interview.
Reliability and scope
Each interview is recorded, with handling for interrupted and stalled recordings so a failure is detected rather than silent.
AI models sit behind one interface, so changing a model is a configuration change.
Concurrent sessions were load-tested so limits are known before they're hit in a real interview.
- Integrity flags are signals, not verdicts. Suspicious behaviour is flagged in the report for the recruiter to judge.
- Evaluation supports a person. The report summarises and structures the interview; it doesn't make the hiring decision.