🚀 C&D Assist - AI-Powered Bug Detection & Fixing

C&D Assist (Code and Debug AI Assistent) is a AI tool that performs automated bug detection and fixing. Designed with a focus on real-time assistance, it integrates a custom-built web-based IDE to provide a seamless debugging experience for developers.

---

✅ Current Implementation

🔧 What It Does

• Takes raw code as input through a browser-based IDE interface.
• Automatically detects bugs and suggests corrected code.
• Allows users to view error outputs, compare fixed code, and make corrections directly inside the IDE.
• Supports uploading Python files, language selection, and custom prompts to control bug-fixing behavior.

⚙️ Key Components

• Frontend:

  • Built using HTML, CSS, JavaScript, and PHP.
  • Integrated with ACE Code Editor for a professional IDE feel.
  • Real-time user interface for input, output, and error visualization.

• Backend:

  • Written in Python.
  • Interacts with transformer-based code models for bug fixing.
  • Receives code input from frontend and returns fixed code or errors.

🤖 Final Model Used

• Model: stabilityai/stablecode-completion-alpha-3b-4k
• Reason for selection:
  • Balanced performance and resource efficiency.
  • Produced the most accurate and usable outputs within hardware limits.
• Minimum System Specs:
  • Intel i5 11th Gen
  • RTX 2050 (4GB VRAM)
  • 16GB RAM

---

🕰️ What We Originally Planned

🧠 Initial Approach

• Used the Defectors Dataset (~150k Python samples).
• Built a classification model using:
  • RandomForestClassifier
  • XGBoost
  • Achieved ~70% accuracy on detecting buggy vs clean code.

🔨 Bug Fixing Plan

• Fixing bugs using DeepSeek-Coder 6.7B.
• Integrated static analysis tools like PyFlakes and AutoPEP8 for basic checks.

❌ Why It Didn't Work

• DeepSeek 6.7B was too large to run on our system.
• PyFlakes/AutoPEP8 lacked depth in identifying logical or structural issues.
• Model-based fixing was necessary — so we pivoted.

---

🔄 Second Attempt: New Datasets, New Models

📚 Datasets Used

• CodeXGLUE (Python subset)
• many-types-4-py-dataset
• QuixBugs

❌ What Went Wrong

• After data cleaning and filtering, dataset was too small for model training.
• Tried to combine datasets, but still:
  • Poor generalization
  • Inconsistent labeling
  • Model couldn’t classify buggy code reliably

---

🧪 Final Attempts Before Deadline

🧠 Tried Generative Models

• deepseek/deepseek-coder-1.3b-instruct → Poor quality responses
• deepseek/deepseek-coder-1.3b-base → Also inconsistent output
• ✅ Final choice: stabilityai/stablecode-completion-alpha-3b-4k
  • Best compromise between output quality and hardware efficiency.

---

🌟 If We Had More Time...

If time and resources permitted, we would have:

• Fine-tuned smaller generative models on domain-specific Python bug datasets.
• Built a multi-stage pipeline:
  1. Detect
  2. Localize
  3. Fix
• Improved dataset size using data augmentation (mutating clean code with known bug patterns).
• Created a feedback loop from user corrections to continuously fine-tune the model.
• Built plugins for VS Code or Jupyter Notebooks for smoother developer experience.

---

📁 Project Structure

C_D_Assit/
├── backend/
│   ├── app.py
│   ├── requirements.txt
│   └── bugfix_model/
│       └── fix_code.py

├── frontend/
│   ├── index.html
│   ├── style.css
│   └── script.js

├── src/
│   ├── api_server.py
│   └── bug_fixer.py

├── models/
│   ├── models--stabilityai--stablecode-completion-alpha-3b-4k/
│   └── stored_tokens/

---

🧑‍💻 Run Locally

1. Clone the repo

   git clone https://github.com/Yuvanraj-K-S/C_D_Assit.git
   cd C_D_Assit
   

2. Set up Python environment

   pip install -r backend/requirements.txt
   

3. Load Model

   python src/bug_fixer.py
   

4. Launch backend API

   python src/api_server.py
   

5. Run Frontend

   Frontend/ui/index.html
   

---