InterviewPro: Agentic AI Mock Interviewer

Problem & Motivation

Millions of job candidates face high-stakes interviews without adequate preparation—not due to lack of effort, but due to limited access to quality resources. Traditional mock interviews with human coaches are expensive, inconsistent, and difficult to scale. As a result, candidates struggle with:

• Lack of access to expert feedback
• No way to practice on demand
• Inability to track improvement over time
• Inconsistent evaluation standards

InterviewPro was built to bridge this interview readiness gap by providing an always-available, scalable, and objective mock interview system. 

Data Source & Data Science Approach

Data Source

The system leverages the RecruitView dataset, a multimodal behavioral dataset designed for interview performance research. It includes:

• 2,011 video responses
• 300+ participants
• 76 curated interview questions
• Multimodal signals: video, audio, and transcripts

This dataset provides realistic, diverse, and human-labeled benchmarks for evaluating interview performance. 

Data Science & Modeling Approach

InterviewPro integrates multiple AI components into an end-to-end pipeline:

1. Resume Parsing

• Extracts structured information (projects, roles, experience) from PDFs/DOCX
• Enables personalized question generation

2. Interview Planning (LLM)

• Uses Amazon Nova Lite to generate tailored behavioral and resume-based questions
• Produces a structured interview plan

3. Real-Time Interview Agent

• Uses OpenAI Realtime for the live voice interview experience
• Conducts live interviews with dynamic follow-up questions
• Maintains conversational flow via tool-based orchestration

5. LLM-Based Scoring System

• Uses a rubric-driven prompt to evaluate responses across dimensions:
  • Answer quality
  • Speaking skills
  • Confidence
  • STAR methodology adherence
• Outputs structured JSON with:
  • Scores
  • Strengths
  • Weaknesses
  • Actionable suggestions

The system is designed with anti-score-inflation prompts to ensure consistent and fair evaluation. 

Evaluation

The model’s performance was validated against human-labeled RecruitView benchmarks:

• 72% of scores fall within ±1.0 of ground truth
• Mean difference: 0.77 (on a 0–10 scale)
• Correlation (Pearson r): 0.40
• Evaluation size: 200 samples (0 failures)

These results demonstrate strong alignment with human evaluation, validating the effectiveness of the LLM-based scoring rubric. 

Key Learnings & Impact

Key Learnings

• Prompt design is critical: Structured rubrics significantly improve scoring consistency
• Agentic systems enable realism: Dynamic questioning improves interview authenticity
• Strict scoring constraints reduce bias and inflation

Impact

• Enables 24/7 interview practice with no scheduling constraints
• Provides objective, standardized feedback at scale
• Empowers users to track improvement over time
• Democratizes access to high-quality interview preparation

Overall, InterviewPro transforms interview preparation from a scarce, human-limited resource into a scalable, AI-driven experience. 

Acknowledgements

• Course: DATASCI 210 Capstone, UC Berkeley
• Instructor: Frederick Nugen, Ramesh Sarukkai
• Team: Hamza Islam, Kyle Ruan, Rui Sun

• Technologies & Platforms:

  • Amazon Nova Lite via AWS Bedrock for interview planning and feedback generation
  • OpenAI Realtime for the live voice interview runtime
  • WebRTC for low-latency live audio interaction
  • RecruitView dataset for benchmarking and evaluation analysis