BEE — Behavioral Evaluation Engine

A mock interview tool that generates questions using a quantized LLM (Qwen2.5-7B via HuggingFace), grades your answers, and gives you a breakdown.

---

What it does

You drop in your skills (or upload a resume), and BEE puts together a 15-question interview split across four sections — theory, aptitude, coding, and HR. Each answer gets evaluated for correctness, depth, and clarity. At the end you get a score, section-wise breakdown, weak areas, and learning resources.

What it can do:

• Extract skills from a PDF or TXT resume automatically
• Generate tailored questions based on your skill set
• Evaluate free-text and code answers
• Rephrase a question if you don't understand it (2 tries per question)
• Show a full review of every question + answer + feedback after the session
• Restart the interview with the same skills in one click

---

Tech Stack

Layer	Has
Backend	Python, FastAPI, Uvicorn
LLM	Qwen2.5-7B-Instruct via HuggingFace Inference API
Resume parsing	PyPDF2
Frontend	Vanilla HTML / CSS / JS
Code editor	CodeMirror 5 (Dracula theme)
Env management	python-dotenv

---

File Structure

bee/
│
├── backend/
│   ├── main.py                  # FastAPI app, all routes
│   ├── interview_controller.py  # Session logic, question flow
│   ├── qwen_client.py           # HuggingFace API calls (generate, eval, rephrase)
│   ├── scoring.py               # Score calculation and verdict logic
│   ├── local_utils.py           # Skill validation, gibberish checks (no API)
│   ├── resources.py             # Static learning resource map
│   └── requirements.txt
│
└── frontend/
    ├── index.html               # Landing page
    ├── interview.html           # Interview page
    ├── results.html             # Results page
    ├── style.css                # All styles (single file)
    ├── landing.js               # Landing page logic
    ├── interview.js             # Interview page logic, CodeMirror setup
    └── results.js               # Results rendering and score animation

main.py auto-detects the frontend folder — it looks for frontend/ or front/ relative to itself.

---

Setup

0. Install Python

This project was built and tested on Python 3.11.9. It's recommended to use the same version to avoid compatibility issues.

Download it from python.org/downloads.

1. Clone the repo

git clone https://github.com/GitRzh/bee.git
cd bee

2. Create and activate a virtual environment

python -m venv venv

# Windows
venv\Scripts\activate

# Mac / Linux
source venv/bin/activate

3. Install dependencies

pip install -r backend/requirements.txt

4. Create a .env file in the backend/ folder

HF_API_KEY=your_huggingface_api_key_here

Get a free key at huggingface.co/settings/tokens. A free account works, but the free tier has rate limits.

5. Run the server

cd backend
python main.py

Open http://localhost:8000 in your browser. That's it. :P

---

Caution :O

A few things to know before you go in:

• HuggingFace free tier runs out. If your monthly credits are gone, question generation and evaluation will fail silently — the app falls back to a local question bank, but grading will stop working. Keep an eye on your HF usage.

• Cold start lag. The first request of the day can take 30–60 seconds because the model has to load on HF's servers. There's a 90-second timeout built in, but if it hits that, just try again.

• The model can hallucinate scores. Qwen grades answers by parsing JSON from an LLM output. If the model returns something malformed, it falls back to a heuristic scorer (based on word count). The heuristic is rough — don't trust a suspiciously high or low score on a long answer.

• Aptitude grading is strict. Math-style answers with symbols and working steps can confuse the model. It's been prompted to handle this, but edge cases exist.

• Sessions are in-memory only. Restarting the server wipes all active sessions. No persistence anywhere.

• Resume parsing is basic. PyPDF2 doesn't handle heavily formatted or scanned PDFs well. If skill extraction looks wrong, use the manual skill entry instead.

---

A project, don't use this as a substitute for actual interview prep.