AI-Powered Code Review and Bug Prediction System

Final Year Project — Institute of Management Sciences (IMSciences), Peshawar
Authors: Muhammad Nasir Khan · Ahmed Khan Khisro · Umair Dost
Supervisor: Mr. Omar Bin Samin
Session: 2021–2025

---

Overview

This project presents a hybrid AI system that automates software code review and predicts bug-prone modules. It combines a Random Forest classifier (trained on 87 software metrics) with an LSTM neural network (operating on tokenised source code) in a soft-voting ensemble, achieving 87.1% accuracy and AUC-ROC of 0.94 on a held-out test set of 2,812 labelled modules.

The system generates structured, actionable review suggestions with SHAP-based metric explanations and exposes its functionality through a FastAPI REST backend and an Angular 16 web dashboard.

---

Key Results

Model	Accuracy	Precision	Recall	F1-Score	AUC-ROC
Random Forest	83.4%	0.81	0.86	0.83	0.91
LSTM	81.7%	0.79	0.88	0.83	0.90
Ensemble (Ours)	87.1%	0.84	0.89	0.87	0.94

Compared to baselines:

System	Accuracy	F1-Score	AUC-ROC
SonarQube (rule-based)	71.2%	0.68	N/A
PMD (rule-based)	68.5%	0.65	N/A
Logistic Regression	76.3%	0.74	0.82
Proposed Ensemble	87.1%	0.87	0.94

---

Architecture

┌─────────────────────────────────────────────────────────┐
│                    Angular Frontend                      │
│         Code Submission · Results Dashboard · Reports    │
└────────────────────────┬────────────────────────────────┘
                         │ HTTP REST
┌────────────────────────▼────────────────────────────────┐
│                  FastAPI Backend                         │
│   /predict/metrics   ·   /predict/code   ·   /health    │
└──────────┬──────────────────────────┬───────────────────┘
           │                          │
  ┌────────▼────────┐       ┌─────────▼────────┐
  │  Random Forest  │       │   LSTM Network   │
  │  (61 metrics)   │       │ (token sequences) │
  └────────┬────────┘       └─────────┬────────┘
           │                          │
           └──────────┬───────────────┘
                      │
             ┌────────▼────────┐
             │ Soft Voting      │
             │ Ensemble (0.55/  │
             │ 0.45 weights)    │
             └────────┬────────┘
                      │
             ┌────────▼────────┐
             │ SHAP Explainer  │
             │ + Review Report │
             └─────────────────┘

---

Repository Structure

ai-code-review/
├── backend/                    # FastAPI REST API
│   ├── app/
│   │   ├── api/                # Route handlers
│   │   ├── models/             # Pydantic schemas
│   │   ├── services/           # ML inference logic
│   │   └── utils/              # Helpers (metrics extractor, tokenizer)
│   ├── tests/                  # Pytest unit & integration tests
│   └── requirements.txt
├── frontend/                   # Angular 16 dashboard
│   └── src/
├── ml/                         # Training & evaluation scripts
│   ├── notebooks/              # Jupyter EDA & training notebooks
│   └── scripts/                # CLI training scripts
├── data/
│   ├── raw/                    # GitHub mining output (gitignored)
│   └── processed/              # Feature matrices (gitignored)
├── configs/                    # Hyperparameter YAML configs
├── docs/                       # Thesis PDF and architecture diagrams
├── .github/workflows/          # CI/CD pipeline (GitHub Actions)
├── docker-compose.yml
└── README.md

---

Quick Start

Prerequisites

• Python 3.10+
• Node.js 18+
• Docker & Docker Compose (optional)

1. Clone the Repository

git clone https://github.com/<your-username>/ai-code-review.git
cd ai-code-review

2. Backend Setup

cd backend
python -m venv venv
source venv/bin/activate          # Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn app.main:app --reload --port 8000

API docs available at: http://localhost:8000/docs

3. Frontend Setup

cd frontend
npm install
ng serve

Dashboard available at: http://localhost:4200

4. Docker (Full Stack)

docker-compose up --build

---

API Reference

POST /predict/metrics

Accepts a JSON payload of extracted software metrics and returns a bug probability score.

Request:

{
  "loc": 245,
  "cyclomatic_complexity": 18,
  "cbo": 12,
  "wmc": 24,
  "lcom": 0.73,
  "rfc": 45,
  "fan_in": 6,
  "fan_out": 9,
  "num_authors": 4,
  "code_churn": 312
}

Response:

{
  "bug_probability": 0.87,
  "prediction": "BUGGY",
  "severity": "HIGH",
  "top_features": [
    {"feature": "cbo", "shap_value": 0.34},
    {"feature": "wmc", "shap_value": 0.28},
    {"feature": "loc", "shap_value": 0.19}
  ]
}

POST /predict/code

Accepts raw source code and returns LSTM-derived code smell classifications and ensemble prediction.

Request:

{
  "source_code": "def process_data(...):\n    ...",
  "language": "python"
}

Response:

{
  "bug_probability": 0.91,
  "prediction": "BUGGY",
  "severity": "HIGH",
  "code_smells": ["God Class", "Long Method", "Duplicate Code"],
  "review_suggestion": "This module exhibits high coupling (CBO = 18) and excessive method complexity (average WMC = 24). Refactoring is recommended: decompose into smaller single-responsibility classes and reduce inter-module dependencies."
}

---

ML Pipeline

Data Collection

• 512 open-source GitHub repositories (Python & Java)
• Selection criteria: ≥500 commits, issue tracker with bug labels, ≥1,000 LOC
• Bug labelling via SZZ algorithm on commit messages

Feature Engineering

• 87 raw metrics extracted per module → 61 features after correlation pruning
• Tools: Radon (Python), CKJMExtended (Java)
• Class imbalance handled with SMOTE (training split only)

Models

Component	Details
Random Forest	200 estimators, balanced class weight, 5-fold CV grid search
LSTM	Embedding(10k vocab, 128-dim) → LSTM(128) → LSTM(64) → Dense(64) → Sigmoid
Ensemble	Soft voting: RF×0.55 + LSTM×0.45, threshold=0.5

To Train from Scratch

cd ml/scripts
python mine_repositories.py --config ../../configs/mining_config.yaml
python extract_features.py --input ../../data/raw --output ../../data/processed
python train_random_forest.py --config ../../configs/rf_config.yaml
python train_lstm.py --config ../../configs/lstm_config.yaml
python evaluate_ensemble.py

---

Testing

cd backend
pytest tests/ -v --cov=app --cov-report=html

---

Citation

If you use this work, please cite:

@misc{khan2025aireview,
  title     = {AI-Powered Code Review and Bug Prediction System},
  author    = {Muhammad Nasir Khan and Ahmed Khan Khisro and Umair Dost},
  year      = {2025},
  school    = {Institute of Management Sciences, Peshawar},
  note      = {BCS (Hons.) Final Year Project}
}

---

License

This project is licensed under the MIT License. See LICENSE for details.