Skip to content
All projects
PrototypeAI Agents

AI Interview System

A multi-agent interview platform design covering interview planning, question generation, answer evaluation, behavioral analysis and report generation.

Overview

A prototype of an AI-assisted interview platform in which specialised agents each own one stage of the interview process, from planning through to a final assessment report.

Problem

Structured interviews are time-consuming to prepare, hard to score consistently, and difficult to summarise into a comparable report.

Solution

Split the process across cooperating agents: a planner defines the interview structure, a question generator tailors questions to the role and profile, an evaluator scores answers, a behavioral analyzer processes visual signals, and a report generator produces the final assessment. Conversational and avatar interfaces are being explored for the candidate experience.

Architecture

AI Interview System architecture, top to bottom
  1. 01

    Interface

    • Candidate UI
    • Conversational / avatar layer
  2. 02

    Backend

    • FastAPI
    • Session orchestration
  3. 03

    Agents

    • Interview planner
    • Question generator
    • Answer evaluator
    • Behavioral analyzer
    • Report generator
  4. 04

    Models & context

    • LLM
    • RAG context
    • MediaPipe / ONNX Runtime
  • One agent per responsibility, coordinated through a FastAPI backend.
  • RAG grounds question generation and evaluation in role and profile context.
  • Behavioral analysis runs on facial-signal models (MediaPipe, ONNX Runtime, HSEmotion).
  • Avatar interfaces (for example Simli) are treated as pluggable presentation layers.

Key features

  • Role-aware interview planning
  • Question generation from role and profile context
  • Answer evaluation and scoring
  • Behavioral / facial-expression signal analysis
  • Automated interview report generation

Engineering challenges

  • Keeping answer evaluation consistent and explainable across candidates and roles.
  • Treating behavioral signals as supporting information rather than definitive judgement.
  • Coordinating several agents while keeping latency acceptable for a live conversation.

Results

No measured outcomes are published for this project. I only list results that have been verified.