← All work
Case study

InterVox

A voice-based mock interview system. Resume and job description in, a spoken interview with per-answer scoring, and a final report with coaching notes.

View source on GitHub →
Role
Creator
Company
Personal project
Stack
LangGraph · Flask · WebSocket · STT / TTS
5agents under one supervisor
Livevoice turns over WebSocket
2APIs: REST and WebSocket
The supervisor graph · five agents
Five agents, one supervisor
  1. PlannerResume + JD become an interview plan
  2. InterviewerTurn-by-turn voice
  3. RouterDecides what happens next
  4. EvaluatorScores every answer
  5. ReportFinal report with coaching notes

The problem

Interview practice is only useful if it feels like an interview: spoken, specific to the role, and followed by honest feedback. InterVox runs the whole loop from a resume and a job description.

What I built

A five-agent supervisor system

A LangGraph supervisor coordinates five agents: planner, interviewer, router, evaluator and report. The planner turns the resume and job description into an interview plan, and the rest run it.

A live voice interview

A Flask backend exposes REST and WebSocket APIs and handles real-time voice turns with speech-to-text and text-to-speech, PDF extraction and session state.

Scoring and coaching

Every answer is scored, and the interview ends with a final report that includes coaching notes.

The calls

Call 1 · Framework

LangGraph here, not in the law agent.

This system has multiple roles, persistent state and branching, which is what a graph is for. The law agent has one decision point per turn, so it doesn’t need one. Right-sizing the tool to the problem.

Call 2 · Transport

WebSocket for the voice turns.

Voice turns have to feel live, so the interview runs over a WebSocket while the rest of the product sits behind REST.