interviewer-gpt

Guru.AI — a GPT-powered mock-interview assistant for ML Engineering, Leadership/Managerial, and Coding interviews. It generates a random interview question, lets you answer by typing (or by recording audio that is transcribed locally with Whisper), and then evaluates your answer with an LLM. For leadership/behavioural questions it also scores how closely your answer maps to the Amazon Leadership Principles using embedding similarity.

The app is built with Gradio and LangChain on top of the OpenAI API.

Repository structure

• interviewer_app.py — the Gradio app (entry point).
• utility/
  • helpers.py — database, embedding/similarity, summarization, transcript and threading helpers.
  • prompts.py — the prompt templates used for each question type.
• resources/
  • interview_questions.db — SQLite database that the question tables are loaded into.
  • sample_behavioral_questions.txt — behavioural/leadership questions (one per line). Add more lines to extend the pool.
  • sample_coding_questions.txt — coding questions (one per line).
  • sample_ML_System_Design_questions.txt — ML system-design questions (one per line).
  • leadership_principles.json — the Amazon Leadership Principles used to evaluate behavioural answers.
  • random_story.txt — an unrelated story used as a similarity baseline.
  • interview_youtube.txt — a sample transcript used by the test.
• tests/
  • test_youtube.py — an end-to-end script that pulls a YouTube mock-interview transcript and runs the evaluation pipeline against it.
• .env.example — template for the local configuration (copy to .env).

Note: the question text files seed interview_questions.db automatically the first time a question table is empty. The database ships pre-seeded and deduplicated, so to pick up questions you add to the .txt files, delete resources/interview_questions.db and run the app again (see Adding your own questions).

Getting started

This project targets Python 3.10–3.12 (verified on 3.12). Some pinned dependencies do not yet ship wheels for Python 3.14.

1. Clone the repository.

2. Create and activate a virtual environment:

   python3.12 -m venv .venv
   source .venv/bin/activate        # Windows: .venv\Scripts\activate

3. Install the dependencies:

   pip install -r requirements.txt

4. Create your local config from the template and add your OpenAI API key:

   cp .env.example .env
   # then edit .env and set OPENAI_API_KEY=sk-...

5. Run the app:

   python interviewer_app.py

6. Open http://127.0.0.1:7860 in your browser.

   To expose a temporary public share link, set GRADIO_SHARE=1 before launching.

Configuration

All settings are read from a local, git-ignored .env file (see .env.example):

Variable	Purpose	Default
OPENAI_API_KEY	Your OpenAI API key (required for any LLM call)	—
MODEL	OpenAI chat model used for evaluation/summarization	gpt-3.5-turbo
TEMPERATURE	Sampling temperature	0.0
MAX_TOKENS	Max tokens for generations	3000
CHUNK_SIZE	Char threshold/size for splitting long responses	600
CHUNK_OVERLAP	Overlap between text chunks when summarizing	30

Adding your own questions

Each question type is backed by a plain-text file in resources/ (one question per line):

• Coding → sample_coding_questions.txt
• Leadership/Behavioural → sample_behavioral_questions.txt
• ML System Design → sample_ML_System_Design_questions.txt

Add your questions on new lines, then rebuild the database so the new entries are picked up:

rm resources/interview_questions.db
python interviewer_app.py   # the tables are re-seeded from the .txt files on first use

Questions are inserted with parameterized queries, so apostrophes and quotes are handled safely.

How it works

1. Choose the question type (Leadership & Behavioural, ML System Design, or Coding).
2. Click Generate me a random interview question.
3. Type your answer — or, for behavioural questions, record audio and it will be transcribed locally with the openai/whisper-base.en model. (The model downloads on first use.)
4. Click Guru evaluation of the response to get:
   • a high-level ranking (Weak / Average / Good / Excellent) with an explanation,
   • for behavioural questions, a per-principle breakdown against the Amazon Leadership Principles plus a similarity bar chart, and
   • a sample "Guru" answer for reference.

Running the test

tests/test_youtube.py is an end-to-end script (not a unit test). It downloads a YouTube transcript, summarizes it, and runs the behavioural evaluation, writing results into the tests/ directory. It requires a valid OPENAI_API_KEY and network access:

python tests/test_youtube.py

Security note

Never commit secrets. The .env file is git-ignored. An earlier version of this repository committed an OpenAI key in resources/conf.env; that file is no longer tracked and any key it contained should be considered compromised and revoked at https://platform.openai.com/api-keys.

Maintenance notes

The codebase was modernized to current libraries and verified end-to-end (UI build, server boot, live OpenAI call, and the full evaluation path):

• LangChain upgraded from 0.0.327 to the 1.x split packages (langchain 1.x, langchain-core, langchain-openai, langchain-text-splitters); calls use .invoke() instead of the removed .predict() / llm(...) / .run() APIs, and the removed load_summarize_chain was replaced with a small manual map-reduce summarizer.
• Gradio upgraded from 4.0.2 to 6.x (gr.update, Audio(sources=...), horizontal BarPlot via x/y + sort instead of the removed vertical=, removed .style() / launch(width/height) / show_download_button); the server launch is guarded under if __name__ == "__main__".
• youtube-transcript-api updated to the v1 instance API (YouTubeTranscriptApi().fetch(...)).
• Chat models and the Whisper transcriber are built lazily, so the UI starts without an API key or a model download.
• Configuration moved from resources/conf.env to a git-ignored .env.
• Fixed a database bug where every question request re-appended the whole .txt file (the table is now seeded only when empty), and switched inserts to parameterized queries.