InterviewPro
Agentic AI Mock Interviewer
DATASCI 210: CAPSTONE · SPRING 2026 · LEC 004
Team: Hamza Islam, Kyle Ruan, Rui Sun
UC Berkeley
The Problem
Millions of candidates enter high-stakes interviews underprepared — not from lack of effort, but lack of access.
Our Solution
No Access
Quality mock interviews and expert feedback remain scarce and expensive
Not Scalable
Human coaching is time-consuming, inconsistent, and out of reach for most
No Tracking
Candidates can't practice on-demand or measure real improvement over time
UC Berkeley
Motivation
An AI-powered interview platform that conducts realistic mock interviews tailored to your resume and target role, then delivers scored, actionable feedback on demand.
Available 24/7
Practice anytime, anywhere
Unlimited Scale
No bottlenecks, no scheduling
Progress Tracking & Objective Feedback
Measure progress with scored reports
Consistent, standardized scoring
Live Demo
UC Berkeley
Who We Surveyed
What We Measured
Key Product Insights
User Experience Feedback
Conversion Signals
Selected feedback:
UC Berkeley
User Feedback
The RecruitView dataset is a multimodal behavioral dataset built for research into personality traits and interview performance using real interview-style responses — providing the grounding needed for robust, realistic scoring.
1
Multimodal by Design
Each sample includes video, audio, and text transcripts — enabling models to learn from visual cues, speech patterns, and content simultaneously
2
Real-World Conditions
Collected in natural environments such as homes and classrooms, maximizing real-world robustness and generalizability
3
Scale & Diversity
2,011 video responses from 300+ unique participants answering 76 curated interview questions
UC Berkeley
Dataset Overview
The InterviewPro platform integrates multiple intelligent components to deliver a seamless and insightful mock interview experience. This architecture is designed for real-time processing and personalized feedback, from resume and job listing parsing to a final actionable feedback report.
Web UI
User uploads resume and a job listing, runs live interviews via mic, and views detailed feedback reports
3. Interview Agent
Orchestrates the live session, adapts follow-up questions in real time.
OpenAI Realtime API + WebRTC
1. Resume/Job Listing Parser
2. Interview Planner
Generates a tailored interview blueprint the agent can reference during the live session.
Amazon Nova 2 Lite
4. Feedback Report
Per-question scorecards with strengths, weaknesses, and actionable improvements. Amazon Nova 2 Lite
UC Berkeley
Solution Overview
Extracts skills, roles, and experience from resumes; incorporates job listing requirements for role-specific question generation.
+
UC Berkeley
Purpose: Generates a personalized interview blueprint from the candidate's resume and target job listing. The planner makes the interview feel realistic and targeted instead of generic, and it gives the live agent a grounded structure to follow during the session.
Model: Amazon Nova 2 Lite via AWS Bedrock
Interview Planner Overview
Inputs:
Output:
UC Berkeley
Purpose: Runs the live mock interview as an autonomous voice agent.
Model: OpenAI Realtime API (gpt-realtime) | Transcription: gpt-4o-transcribe | Transport: WebRTC
1
Prompt Engineering
The interview plan is embedded in the system prompt allowing the agent to autonomously adapt question order and phrasing based on the conversation.
2
Agent Tools
The Agent can call end_interview to request wrap-up when the plan is complete. Backend runs in parallel to track questions, transcribe audio, and prepare for the feedback report.
3
Audio Transport via WebRTC
Browser captures candidate audio via mic and streams directly to the Agent for low-latency, peer-to-peer audio.
Interview Agent Overview
Per-Question Scorecards
Individual ratings for every response in the session
Top Strengths
Highlights what the candidate communicated well
Top Weaknesses
Pinpoints gaps in structure, depth, or delivery
Actionable Feedback
Specific, improvement-focused suggestions aligned with RecruitView research data
Powered by Amazon Nova Lite, each feedback report gives candidates a structured, research-aligned view of their performance.
UC Berkeley
Feedback Report Overview
Model & Approach
Baseline evaluation uses Amazon Bedrock (amazon.nova-lite-v1) with a carefully engineered prompt to score candidate responses consistently and fairly.
Structured JSON Output
Scores
Per-dimension numeric ratings
Strengths
Identified positive signals
Weaknesses
Gaps and structural issues
Suggestions
Specific improvement actions
1
Multi-Dimensional Grading
Scores answer quality, speaking skills, confidence, and overall performance using a strict rubric
2
STAR Score Assignment
Assigns a structured STAR score to behavioral responses, penalizing vague or poorly structured answers
3
Anti-Inflation Design
Prompt explicitly avoids score inflation and enforces strict scoring ranges for consistency
UC Berkeley
Feedback Report/LLM-Based Rubric Transcript Scoring
72% of Nova scores fall within ±1.0 point of dataset ground truth across 200 evaluated samples — demonstrating strong alignment between the LLM rubric and human-validated benchmarks.
72%
Within ±1.0 Pts
Answer score alignment rate across 200 samples
0.77
Mean Difference
Overall average score deviation (0–10 scale)
0.40
Pearson r
Overall correlation between Nova and dataset scores
200
Samples Evaluated
Full evaluation run
UC Berkeley
Feedback Report Model Performance
Nova Evaluation Snapshot
Question: "What has been your greatest failure?"
✅ Strength: Clear, specific failure with demonstrated resilience ⚠️ Improve: Deeper reflection on learning and measurable outcomes
Answer Quality
4.5 / 10
Speaking Skills
5.0 / 10
Confidence
4.0 / 10
STAR Score
3 / 5
UC Berkeley
Feedback Report Model Output — Good Score Example
Nova Evaluation Snapshot
Question: "Where do you see yourself in the next 5 years?"
⚠️ Improve:
Answer Quality
2.5/ 10
Speaking Skills
2.5/ 10
Confidence
2.0/ 10
STAR Score
1 / 5
UC Berkeley
Feedback Report Model Output — Bad Score Example
UC Berkeley
CONCERNS
Privacy Risks
Resume, audio, and video contain sensitive personal data
Bias
Scoring algorithms result in unfair treatment across demographics and communication styles
Data Transparency
Unclear usage policies erode user trust
Data Minimization
Risk of collecting more data than necessary causes concern
MITIGATIONS
Security
End-to-end encryption, authentication, and strict access controls
Bias Reduction
Research-based, objective prompts with minimal ambiguity and minimal language barriers
Transparency
Clear data usage disclosures and deletion options, seeking consent before collecting users’ data
Minimization
Collecting only what is essential; delivering personalized value in return
Ethics and Privacy
UC Berkeley
Our roadmap includes immediate enhancements and strategic pillars that will redefine the mock interview experience.
Multimodal Analysis
Analyzing body language, facial expressions, and tone
Personalized Learning Paths
Customizing coaching based on individual strengths, weaknesses, mock interview scores, and career goals.
Broader Integration
Seamlessly connecting with applicant tracking systems and other career development platforms.
Language Assistance
Providing language support for users with limited English proficiency, such as translation of interview questions
Future Improvements
UC Berkeley
Acknowledgements
Instructor: Frederick Nugen, Ramesh Sarukkai
Technologies & Platforms:
Thank you!
Hamza: Resume/Job Listing Parser, Interview planner, Interview agent, Web UI, AWS deployment
Rui: Data Gathering, Model Training, Feedback Report/Scoring, User Feedback
Kyle: User Feedback, Interview Model Exploration,, Ethics and Privacy, Future Improvements
Contributions