Qais Ghazal | AI & Machine Learning Engineer

About Me

My Journey in AI

The best way to predict the future is to create it

I am a final-year student of Data Science and Artificial Intelligence at Al Hussein Technical University, passionate about leveraging AI to create meaningful solutions for complex problems.

My research and projects focus on computer vision, natural language processing, and the intersection of machine learning with practical applications.

Download Resume

Live Demo

NLP

MedClassify - AI Medical Assistant

Production-ready Arabic medical AI that helps patients find the right specialist. Built with fine-tuned MARBERT achieving 90%+ accuracy on 90K+ medical consultations, deployed as a responsive bilingual web app.

MARBERT Flask Arabic NLP Healthcare AI

Live Demo

Computer Vision

Brain Tumor MRI Classification

Deployed deep learning system for classifying brain tumors from MRI scans into four categories with 99.24% accuracy using EfficientNet-B3 and comprehensive model comparison.

Python PyTorch Flask Medical AI

Live Website

Computer Vision • NLP

DeepKnight - Interactive AI Chess Board

Advanced robotic chess system integrating autonomous movement, real-time computer vision, and adaptive AI. Features emotion detection, voice commands, and 99.7% positioning accuracy with sub-millimeter precision.

Computer Vision NLP Robotics AI/ML

Live Demo

Computer Vision

Image Deep Fake Detection

A CNN-based model that detects deep fake images with 97.4% accuracy using transfer learning on balanced datasets of real and AI-generated images.

Python TensorFlow CNNs Computer Vision

Live Demo

Data Science

Spotify Music Dashboard

Interactive dashboard for exploring Spotify music data, analyzing relationships between audio features and popularity across artists and genres.

R/Shiny Plotly Spotify API dplyr

Data Science

SmartCatFill Framework

Developed a flexible ML framework for preprocessing categorical missing data, achieving 99.43% downstream prediction accuracy versus traditional imputation methods.

Python Scikit-learn Pandas ML Imputation

Computational Neuroscience

FitzHugh-Nagumo Neural Dynamics

Advanced simulation framework modeling neuronal dynamics with parameter optimization based on real electrophysiological data.

Python SciPy Computational Modeling Parameter Optimization

My Expertise

Technical Skills

The technologies and domains I specialize in

Programming

Python
R
Java
SQL
C

Machine Learning

TensorFlow
PyTorch
Scikit-Learn
Neural Networks
Model Deployment

Computer Vision

OpenCV
Convolutional Neural Networks
Image Processing
Object Detection
Face Recognition

Natural Language Processing

NLTK
spaCy
Transformers
Text Classification
Sentiment Analysis

Data Engineering

Pandas
NumPy
SQL Databases
Data Pipelines
ETL Processes

Tools & Platforms

Docker
Git
Hadoop
Spark
Power BI

Background

Education & Experience

My academic journey and professional growth

2021 - Present

Bachelor's in Data Science and Artificial Intelligence

Al Hussein Technical University

Specializing in AI with coursework in machine learning, deep learning, computer vision, natural language processing, and data science. Maintaining a 3.74 GPA

Machine Learning Deep Learning Computer Vision Natural Language Processing

2024

Deep Learning Specialization

Coursera - deeplearning.ai

Completed a comprehensive deep learning program covering neural networks, CNNs, RNNs, and sequence models. Developed practical implementations and earned certification with honors.

Neural Networks CNNs RNNs Deep Learning

2023

Machine Learning Specialization

Stanford University - Coursera

Comprehensive ML program covering supervised and unsupervised learning, recommender systems, and reinforcement learning. Implemented algorithms from scratch and applied them to real-world datasets using Python and scikit-learn.

Supervised Learning Unsupervised Learning Python

Let's Connect

I'm currently looking for internship opportunities in AI and machine learning. Feel free to reach out if you have any questions or would like to collaborate on a project!

Email Me

The fastest way to reach me for professional inquiries

qaisghazal45@gmail.com

Call Me

Available for calls during business hours

+962 777797819

Location

Based in Jordan, available for remote opportunities

Amman, Jordan

Connect With Me

Follow my work and projects on these platforms

Your Name

Your Email

Subject

Message

MedClassify - AI Medical Assistant

A production-ready Arabic medical AI web application that analyzes patient symptoms and recommends the appropriate medical specialty. Using a fine-tuned MARBERT model trained on 25,000+ annotated medical consultations, MedClassify classifies symptoms into 5 specialties: Dentistry, Ophthalmology, Orthopedics, Psychiatry, and Nutrition.

Methodology

The approach combines:

Systematic comparison of 20+ ML approaches across traditional ML, neural networks, and transformers
MARBERT fine-tuning with 4 epochs, learning rate 2e-5, batch size 16
Custom Arabic preprocessing pipeline handling tashkeel removal and normalization
Stratified 70/15/15 train/validation/test split maintaining class balance
Enhanced preprocessing with ISRI stemming and stop word removal

Key Results

MARBERT: 90%+ test accuracy (selected model for deployment)
Traditional ML: 85-89% accuracy with Logistic Regression + TF-IDF
Neural Networks: 83-88% accuracy with GRU/LSTM + embeddings
Enhanced preprocessing improved accuracy by 3-7% across all models
Real-time inference with bilingual Arabic-English web interface

Production Deployment

Responsive web application with native Arabic RTL support
Flask API backend with PyTorch inference pipeline
Smart examples for quick testing across all 5 medical specialties
Privacy-focused design with no data storage
Deployed on Heroku with Docker containerization

Technologies Used

MARBERT

Python

Flask

PyTorch

JavaScript

Arabic NLP

Docker

Try Live Demo View Code

Brain Tumor MRI Classification

A comprehensive deep learning system for automated brain tumor classification from MRI scans. This project implements and compares four state-of-the-art CNN architectures to classify brain tumors into glioma, meningioma, pituitary tumors, and normal tissue with clinical-grade accuracy.

Methodology

The approach combines:

Systematic comparison of four CNN architectures: EfficientNet-B3, DenseNet-121, ResNet-50, and VGG-16
2×2 factorial hyperparameter optimization design with learning rate and dropout tuning
Comprehensive dataset of 7,023 professionally annotated MRI scans
Stratified train/validation/test split (70%/15%/15%) maintaining class balance
Transfer learning with ImageNet pre-trained weights and medical image preprocessing

Key Results

EfficientNet-B3: 99.24% test accuracy (best performer) with optimal efficiency trade-off
DenseNet-121: 97.82% test accuracy with excellent feature reuse capabilities
VGG-16: 90.70% test accuracy demonstrating classical architecture effectiveness
ResNet-50: 87.19% test accuracy with architectural limitations identified
Dataset size analysis showing 96.8% accuracy achievable with just 30% of training data
Real-time inference: 70.6ms CPU, 14.2ms GPU per image with EfficientNet-B3

Clinical Impact

Web-based interface for real-time brain tumor classification
Support for four tumor types: glioma, meningioma, pituitary, and normal tissue
Multi-model comparison capability for enhanced diagnostic confidence
Production deployment with Flask framework and containerized architecture

Technologies Used

Python

PyTorch

Flask

Deep Learning

Medical AI

Docker

GitHub Repository Live Demo

DeepKnight - Interactive AI Chess Board

An advanced interactive AI chess board that seamlessly integrates autonomous robotics, real-time computer vision, and natural language processing. This capstone project demonstrates cutting-edge engineering in human-machine interaction through physical chess gameplay with emotion-aware AI.

Computer Vision Capabilities

Real-time facial emotion recognition using DeepFace neural networks
Automated chess piece detection and board state analysis
2K webcam integration with OpenCV processing pipeline
Continuous behavioral monitoring for adaptive gameplay responses
89.2% accuracy across 7 emotion categories (Happy, Sad, Angry, Fear, Surprise, Disgust, Neutral)

Natural Language Processing

Voice command recognition for chess move input and system control
Conversational AI integration via Gemini API for strategic commentary
Natural language chess move parsing and validation
Real-time speech-to-text processing with contextual understanding
Interactive communication during gameplay sessions

Robotics & AI Integration

Autonomous CoreXY electromagnetic positioning with sub-millimeter precision
Stockfish-powered chess computation with cloud-enhanced processing
Master-slave architecture coordinating Raspberry Pi 4 brain with Arduino Uno motor control
Adaptive learning framework analyzing player behavior patterns and skill progression
Real-time system integration with 98.3% uptime reliability

Performance Metrics

Positioning Accuracy: 99.7% (±0.5mm precision)
AI Response Time: 0.3 seconds average
Emotion Detection: 89.2% accuracy
Voice Recognition: 92.6% accuracy with 3-7s latency
System Success Rate: 97.3% overall reliability
Average Move Cycle: 6.7 seconds complete execution

Technologies Used

Python

OpenCV

DeepFace

Gemini API

Raspberry Pi

Arduino

Stockfish AI

Speech Recognition

Project Website

Image Deep Fake Detection

This project addresses the growing concern of deepfake images by developing a robust detection system using convolutional neural networks and transfer learning. The model achieves 97.40% accuracy by analyzing patterns in real and AI-generated facial images.

Methodology

The approach combines:

Utilization of three CNN architectures: DenseNet121, ResNet50, and InceptionV3
Transfer learning to improve performance with limited training data
Balanced dataset of real images (FFHQ) and fake images (StyleGAN)
Comprehensive evaluation using precision, recall, and F1-score metrics

Key Results

DenseNet121: 97.40% accuracy with 0.067 loss (best performer)
InceptionV3: 97.23% accuracy with 0.073 loss
ResNet50: 85.67% accuracy with 0.438 loss
DenseNet121 achieved balanced precision (0.99 for fake, 0.96 for real) and recall (0.96 for fake, 0.99 for real)

Technologies Used

Python

TensorFlow

CNNs

Computer Vision

Transfer Learning

Try Live Demo GitHub Repository

SmartCatFill: ML-Based Categorical Data Preprocessing

This project introduces a flexible machine learning framework for handling missing categorical data in datasets, resulting in significant improvements in downstream model accuracy compared to traditional imputation methods.

Methodology

Key innovations include:

Flexible algorithm architecture supporting any ML technique for imputation
Intelligent pattern recognition for categorical missing data
Implementation with multiple algorithms (KNN, Decision Trees, Naive Bayes)
Extensible framework designed for future numeric data support

Key Results

99.43% downstream prediction accuracy with Decision Tree implementation
98.21% accuracy with KNN and 84.15% with Naive Bayes implementations
Significantly outperforms traditional categorical imputation methods

Technologies Used

Python

Scikit-learn

Pandas

NumPy

Machine Learning

GitHub Repository

FitzHugh-Nagumo Neural Dynamics Simulation

A sophisticated computational framework implementing the FitzHugh-Nagumo model to simulate neuronal dynamics with parameter optimization using real electrophysiological data from the Allen Brain Atlas.

Technical Approach

Key components include:

Implementation of the FitzHugh-Nagumo differential equations system
Advanced numerical integration methods for stable simulations
Parameter optimization algorithms using empirical neuron recordings
Phase-plane analysis with bifurcation diagrams
Visualization of neuronal responses to various stimuli

Key Achievements

Successfully modeled complex neuronal firing patterns with high accuracy
Optimized differential equation parameters to match real neuronal data
Created comprehensive phase-space visualizations of neural dynamics
Developed an extensible framework for testing various neuronal models
Achieved parameter optimization with < 2% error compared to empirical data

Technologies Used

Python

SciPy

Computational Neuroscience

Matplotlib

Parameter Optimization

Allen Brain Atlas API

GitHub Repository

Spotify Music Dashboard

An interactive data visualization dashboard for exploring Spotify music data, built with R Shiny. This application allows users to analyze relationships between songs' audio features and their popularity.

Methodology

The approach includes:

Integration with Spotify Web API for real-time artist and album information
Data processing and analysis using dplyr and other R packages
Interactive visualizations created with Plotly
Custom styling to match Spotify's visual identity

Key Features

Filter music by artist, album, and various audio features
Interactive visualizations showing relationships between audio features and popularity
View album details and track listings with links to Spotify
Analyze trends in music characteristics across different artists and genres

Technologies Used

R/Shiny

Plotly

Spotify Web API

dplyr

GitHub Repository View Dashboard