Inspiration
Technical interviews lack consistency. Different interviewers ask different questions, assess differently, and struggle to adapt. We wanted to build an AI system that provides fair, structured interviews while feeling genuinely human. The goal was to eliminate bias, scale assessments globally, and create an experience that candidates don't dread—one that actually feels like talking to an experienced interviewer.
What it does
Conducts natural, conversational technical interviews that adapt in real-time. The system dynamically selects questions from a knowledge base based on topic and difficulty, evaluates answers instantly with scoring and feedback, and adjusts question difficulty based on performance. It maintains human-like conversation flow with natural pacing and follow-ups. Supports voice-based interviews through speech integration. Provides structured feedback highlighting strengths and areas for improvement.
How we built it
Backend built with Python and FastAPI for high-performance API handling. Frontend uses HTML and JavaScript for real-time communication. Google ADK orchestrates three specialized AI agents: Interview Agent manages conversation flow, Question Selector Agent retrieves questions from Pinecone RAG knowledge base, and Evaluator Agent scores answers and provides feedback. Pinecone serves as the vector database for semantic question retrieval. Speech integration enables voice-based interviews.
Challenges we ran into
Coordinating multiple agents while maintaining natural conversation flow without disrupting pacing. Balancing real-time evaluation speed with accuracy. Managing conversation context across multiple agent calls. Integrating speech processing to handle natural speech patterns, and pauses. Calibrating dynamic difficulty adjustment so question transitions feel natural. Preventing information loss as context passes between agents. Ensuring the system feels conversational rather than robotic despite complex backend orchestration.
Accomplishments that we're proud of
Built a fully functional multi-agent system that genuinely feels like talking to a real interviewer. Successfully integrated Pinecone RAG for context-aware question selection. Achieved real-time evaluation without sacrificing conversational quality. Implemented working speech integration for immersive voice interviews. Deployed with FastAPI for reliable, scalable performance. The system successfully adapts difficulty based on candidate performance while maintaining engagement.
What we learned
building a multi-agent app in 4 hours is not easy
What's next for Multi-Agent AI Interview Bot
enahnced memory management plagiarism and ai checks additional context using web or other nsources detailed feedback with suggestions beyond thecnical skills
Built With
- fastapi
- gemini
- google-adk
- html5
- javascript
- pinecone
- python
- rag
- webhooks
Log in or sign up for Devpost to join the conversation.