Interview Bot

An AI-powered interview bot designed for conducting technical interviews for AI Engineering positions, with a focus on RAG (Retrieval Augmented Generation) roles. The bot evaluates candidates on Python experience, project examples, RAG understanding, evaluation metrics, and open-source contributions, providing feedback on performance and areas for improvement.

Architecture

The bot comprises three main components:

1. Speech-to-Text (STT): Utilizes Deepgram's Nova model for accurate transcription.
2. Language Model (LLM): Employs Llama 3.2 on Ollama for generating contextual questions and feedback.
3. Text-to-Speech (TTS): Uses Deepgram TTS for natural-sounding voice output.

Prerequisites

• Deepgram API key (for STT and TTS)
• Daily.co API key (for video/audio communication)
• Ollama installed with Llama 3.2 model

Environment Setup

Create a .env file with the following content:

DEEPGRAM_API_KEY=your_key_here
DAILY_API_KEY=your_key_here

Installation

1. Create and activate a virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

2. Install dependencies:

pip install -r requirements.txt

Running the Bot

Start the interview bot:

python interview_bot.py

The bot will conduct a technical interview and provide feedback on the candidate's performance.

Technical Documentation

Components Overview

• Daily Transport Layer: Manages audio/video communication with VAD (Voice Activity Detection) using Silero.

• Speech Services:

  • STT: Deepgram's Nova model for English (en-US).
  • TTS: British male voice profile.

• LLM Service: Llama 3.2 model for question generation and response processing.

System Prompt Structure

The bot follows a structured interview format with:

• Interviewer persona: Allysa from Meta
• Fixed question set covering key topics
• Concludes with personalized feedback

Configuration System

• Environment variables for API keys
• Daily.co room configuration via configure() function
• Supports command-line arguments and environment variables

Voice Activity Detection

Silero VAD with configurable parameters:

• Stop threshold: 0.2 seconds
• Start threshold: 0.2 seconds
• Confidence threshold: 0.4

Call Architecture Flow

Communication Layer

• WebRTC for video call interface, hosted through Daily.co

Signal Processing Flow

1. Input: User's voice captured via WebRTC
2. VAD: Silero VAD manages speech segmentation
3. Processing Pipeline:
   • User Speech → Deepgram STT → Text → Llama 3.2 → Generated Response → Deepgram TTS → Audio Output

This setup ensures a seamless, natural conversation flow, allowing the AI interviewer to engage in real-time dialogue and provide appropriate follow-up questions or feedback.