AI Mock Interviewer

This tool is designed to help you practice various technical interviews by simulating real interview experiences. You can enhance your skills in coding, (machine learning) system design, and other topics. You can brush your interview skills in a realistic setting, although it’s not intended to replace thorough preparations like studying algorithms or practicing coding problems.

Key Features

• Speech-First Interface: Talk to the AI just like you would with a real interviewer. This makes your practice sessions feel more realistic.
• Various AI Models: The tool uses three types of AI models:
  • LLM (Large Language Model): Acts as the interviewer.
  • Speech-to-Text and Text-to-Speech Models: These models help to mimic real conversations by converting spoken words to text and vice versa.
• Model Flexibility: You can use many different models, including those from OpenAI, open-source models from Hugging Face, and locally running models.
• Streaming Mode: All models can be used in streaming mode. Instead of waiting for the full response from the AI, you can get partial responses in real-time.
• DevContainer Support: The project includes a DevContainer configuration for Visual Studio Code and CodeSpaces, making it easy to set up and run the project in a containerized environment.

Running the Simulator

Initial Setup

First, clone the project repository to your local machine with the following commands:

git clone https://github.com/senz/ai-interviewer.git
cd ai-interviewer

Configure the Environment

Create a .env file from the provided Open AI example and edit it to include your OpenAI API key (learn how to get it here: https://platform.openai.com/api-keys):

cp .env.openai.example .env
nano .env  # You can use any text editor

If you want to use any other model, follow the instructions in Models Configuration section.

Using DevContainer

Visual Studio Code

If you have Visual Studio Code installed, you can use the provided DevContainer configuration to set up a development environment with all dependencies pre-installed. Just open the project in VS Code and click on the "Reopen in Container" button.

Once container created, all dependencies will be installed and you can run the application. Make sure to set up the .env file as described above.

poetry run python app.py

VSCode should detect publish port and suggest to open the application in your browser.

Command Line

You can also use devcontainer-cli to open the project in a container:

devcontainer open

CodeSpaces

untested

If you are using GitHub CodeSpaces, you can open the project in a new CodeSpace by clicking on the "Code" button and selecting "Open with CodeSpaces". Provide suggested secrets and you should be good to go. Use port publish/forwarding to access the application.

Running Locally (alternative)

If you don't want to use Docker just set up a Python environment and install dependencies to run the application locally:

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python app.py

The application should now be accessible at http://localhost:7860.

Models Configuration

AI Interviewer is powered by three types of AI models: a Large Language Model (LLM) for simulating interviews, a Speech-to-Text (STT) model for audio processing, and a Text-to-Speech (TTS) model to read LLM responses. You can configure each model separately to tailor the experience based on your preferences and available resources.

Large Language Model (LLM)

• OpenAI Models: You can use models like GPT-3.5-turbo, GPT-4, GPT-4o or others provided by OpenAI. Set up is straightforward with your OpenAI API key.
• Hugging Face Models: Models like Meta-Llama from Hugging Face can also be integrated. Make sure your API key has appropriate permissions.
• Claude: You can use models from Anthropic, such as Claude, for a different interview experience. Ensure you have the necessary API key and permissions.
• Local Models: If you have the capability, you can run models locally using Ollama or other tools. Ensure they are compatible with the Open AI or Hugging Face API for seamless integration.

Speech-to-Text (STT)

• OpenAI Whisper: Available via OpenAI, this model supports multiple languages and dialects. It is also available in an open-source version on Hugging Face, giving you the flexibility to use it either through the OpenAI API or as a locally hosted version.
• Other OS models: Can be used too but can require a specific wrapper to align with API requirements.

Text-to-Speech (TTS)

• OpenAI Models: The "tts-1" model from OpenAI is fast and produces human-like results, making it quite convenient for this use case.
• Other OS models: Can be used too but can require a specific wrapper to align with API requirements. In my experience, OS models sound more robotic than OpenAI models.

Configuration via .env File

The tool uses a .env file for environment configuration. Here’s a breakdown of how this works:

• API Keys: Whether using OpenAI, Hugging Face, or other services, your API key must be specified in the .env file. This key should have the necessary permissions to access the models you intend to use.
• Model URLs and Types: Specify the API endpoint URLs for each model and their type (e.g., OPENAI_API for OpenAI models, HF_API for Hugging Face or local APIs).
• Model Names: Set the specific model name, such as gpt-4o or whisper-1, to tell the application which model to interact with.

Example Configuration

OpenAI LLM:

OPENAI_API_KEY=sk-YOUR_OPENAI_API_KEY
LLM_URL=https://api.openai.com/v1
LLM_TYPE=OPENAI_API
LLM_NAME=gpt-3.5-turbo

Claude LLM:

ANTHROPIC_API_KEY=sk-ant-YOUR_ANTHROPIC_API_KEY
LLM_TYPE=ANTHROPIC_API
LLM_NAME=claude-3-5-sonnet-20240620

Hugging face TTS:

HF_API_KEY=hf_YOUR_HUGGINGFACE_API_KEY
TTS_URL=https://api-inference.huggingface.co/models/facebook/mms-tts-eng
TTS_TYPE=HF_API
TTS_NAME=Facebook-mms-tts-eng

Local STT:

HF_API_KEY=None
STT_URL=http://127.0.0.1:5000/transcribe
STT_TYPE=HF_API
STT_NAME=whisper-base.en

You can configure each models separately. Find more examples in the .env.example files provided.

Running tests

export SKIP_LIVE=1 # Skip tests that require live API access
poetry run pytest

Acknowledgements

Forked from IliaLarchenko/Interviewer

The service is powered by Gradio.

Even though the service can be used with great variety of models I want to specifically acknowledge a few of them:

• OpenAI: For models like GPT, Whisper, and TTS-1. More details on their models and usage policies can be found at OpenAI's website.
• Meta: For the Llama models, particularly the Meta-Llama-3-70B-Instruct, as well as Facebook-mms-tts-eng model. Visit Meta AI for more information.
• HuggingFace: For a wide range of models and APIs that greatly enhance the flexibility of this tool. For specific details on usage, refer to Hugging Face's documentation.

Important Legal and Compliance Information

Acceptance of Terms

By utilizing this project, in any form—hosted or locally run—you acknowledge and consent to the terms outlined herein. Continued use of the service following any modifications to these terms constitutes acceptance of the revised terms.

General User Responsibilities

Users of this project are responsible for complying with all applicable laws and regulations in their jurisdiction, including data protection and privacy laws.

Liability Disclaimer

The creator of this open source software disclaims all liability for any damages or legal issues that arise from the use of this software. Users are solely responsible for their own data and ensuring compliance with all applicable laws and regulations.

License Compatibility

This project is released under the Apache 2.0 license. Users must ensure compatibility with this license when integrating additional software or libraries.

Specific Guidelines for Usage

Running the Service Locally

• Absolute User Responsibility: When the service is run locally, users have absolute control and responsibility over its operation. Users must secure their own API keys from third-party providers or opt to run local models. Users are fully responsible for ensuring that their use complies with all applicable laws and third-party policies.
• Data Sensitivity Caution: Users are strongly cautioned against entering sensitive, personal, or non-public information, including but not limited to trade secrets, undisclosed patents, or insider information that could potentially result in legal repercussions or breaches of confidentiality.

AI-Generated Content Disclaimer

• Nature of AI Content: Content generated by this service is derived from artificial intelligence, utilizing models such as Large Language Models (LLM), Speech-to-Text (STT), Text-to-Speech (TTS), and other models. The service owner assumes no responsibility for the content generated by AI. This content is provided for informational or entertainment purposes only and should not be considered legally binding or factually accurate. AI-generated content does not constitute an agreement or acknowledge any factual statements or obligations.