Overview: A semi-automatic video segmentation tool that accelerates the preparation of image segmentation data for video sequences using GUI-based annotation and AI propagation.

Technologies: Python, PyQt6, PyTorch, OpenCV, OSVOS (One-Shot Video Object Segmentation)

Key Features:

• • Manual annotation tools (pencil, polygon)
• • Semi-automatic label propagation using Optical Flow and OSVOS
• • Frame-by-frame navigation with multi-label support
• • MVC architecture for scalable design

Impact: Combines manual annotation capabilities with automated segmentation algorithms to significantly reduce manual labeling effort for video datasets.

Overview: A comprehensive medical AI system for skin cancer detection using Swin Transformer with parameter-efficient bottleneck adapters. Supports both binary classification (benign vs malignant) and multiclass classification (HAM10000 7-class dataset) for clinical-grade dermatoscopic image analysis.

Technologies: Python, PyTorch, Swin Transformer, Vision Transformer (ViT), Parameter-Efficient Fine-Tuning, Bottleneck Adapters, Medical Imaging

Key Features:

• • Parameter-efficient fine-tuning with bottleneck adapters
• • Swin Transformer and ViT architectures for medical imaging
• • Binary and 7-class classification capabilities
• • HAM10000 dataset integration for dermatoscopic analysis

Impact: Advances dermatological AI by combining efficient transfer learning techniques with state-of-the-art vision transformers for accurate skin cancer detection and classification in clinical settings.

Overview: A real-time facial emotion detection system using YOLOv8 for face detection and FER for emotion analysis across 7 categories.

Technologies: Python, YOLOv8, OpenCV, Facial Expression Recognition (FER), ONNX

Key Features:

• • Real-time emotion recognition across 7 categories
• • Multiple input sources (images, videos, webcam)
• • Two-stage detection pipeline
• • Command-line interface with flexible parameters

Applications: Combines YOLOv8 face detection with emotion recognition algorithms for real-time video streams, suitable for demographic analysis and human-computer interaction applications.

Overview: A comprehensive framework for EEG signal analysis and classification, providing baseline implementations of classical machine learning models for the research community.

Technologies: Python, NumPy, SciPy, scikit-learn, PyTorch, MNE-Python, XGBoost

Key Features:

• • Support for multiple EEG datasets (CHB-MIT, BCI Competition 2a, LEE, Klinik)
• • Classical ML models with signal processing techniques
• • Modular architecture with cross-validation
• • Integration with Weights & Biases for experiment tracking

Impact: Provides standardized framework for EEG signal processing including epilepsy detection and motor imagery classification with reproducible baseline performance metrics across different datasets.

Overview: Signal processing system for detecting and analyzing vibration-induced noise patterns using machine learning techniques and acoustic signal analysis.

Technologies: Python, Signal Processing, Machine Learning, Audio Analysis, Spectral Analysis

Key Features:

• • Vibration pattern recognition and classification
• • Spectral analysis for noise characterization
• • Real-time signal processing capabilities
• • Anomaly detection in mechanical systems

Applications: Enables predictive maintenance and fault detection in mechanical systems by analyzing vibration signatures and identifying abnormal noise patterns for industrial monitoring.

Overview: Learn meaningful EEG representations without labeled data across multiple tasks and datasets using self-supervised learning techniques.

Technologies: PyTorch Lightning, Python, Neural Networks, EEG Signal Processing

Key Features:

• • Self-supervised learning framework for EEG
• • Task-agnostic embeddings for multiple datasets
• • Modular neural architecture with "Minion Networks"
• • Cross-validation support for robust evaluation

Impact: Creates transferable embeddings across different neurological signal analysis tasks using auxiliary tasks like temporal context prediction and channel reconstruction.

Overview: EEG-LTENT (Learned Task-agnostic Embeddings for Neural Time-series) provides a foundation model approach for EEG signal analysis. Instead of training task-specific models from scratch, this framework learns universal EEG representations that can be fine-tuned for various downstream applications.

Technologies: Python, PyTorch, Conformer Architecture, Vector Quantization, Masked Autoencoding, Self-Supervised Learning, CNN-Transformer Hybrid

Key Features:

• • Task-agnostic learning: Pre-train once, adapt to many tasks (classification, regression, clustering)
• • Self-supervised approach using masked autoencoding on unlabeled EEG data
• • Discrete representations through vector quantization for interpretable embedding spaces
• • Universal embeddings that transfer across datasets, tasks, and subjects

Impact: Revolutionizes EEG analysis by creating a foundation model that learns from unlabeled data and transfers knowledge across different neurological tasks, significantly improving performance on limited labeled datasets.

Overview: A comprehensive image captioning system that automatically generates descriptive text for images using deep learning techniques and attention mechanisms.

Technologies: Python, PyTorch, Computer Vision, Natural Language Processing, CNN, RNN, Attention

Key Features:

• • Encoder-decoder architecture with attention mechanism
• • CNN-based image feature extraction
• • RNN-based text generation
• • Multi-modal learning approach

Impact: Bridges computer vision and natural language processing to generate human-like descriptions of visual content for accessibility and automated content description applications.

Overview: A Python application that processes PDF documents and extracts structured information using Large Language Models, designed specifically for medical reports.

Technologies: OpenAI GPT, Meta's LLaMA, Python, PyTorch, Hugging Face Transformers, LangChain

Key Features:

• • Dual LLM support (GPT and LLaMA)
• • Automated PDF content extraction with structured output
• • Batch processing capabilities
• • Resource usage monitoring

Impact: Uses large language models to parse PDF files and extract patient information, test results, and panel summaries with flexible LLM backends and structured CSV/JSON outputs.

Overview: Classify ten classes of Persian text datasets using GloVe embedding to address unique challenges of Persian language processing in multi-class classification.

Technologies: Python, Jupyter Notebook, GloVe Embedding, Machine Learning

Key Features:

• • Multi-class Persian text classification framework
• • GloVe word embedding implementation for Persian
• • Comprehensive preprocessing pipeline
• • Specialized handling of Persian language characteristics

Applications: Leverages GloVe embeddings to convert Persian text into numerical representations for effective machine learning classification across ten distinct text categories.

Overview: A comprehensive machine learning project for tumor subtype classification using mutation and copy number variation data from genomic analysis.

Technologies: Python, Scikit-learn, Pandas, NumPy, Jupyter Notebook, Deep Neural Networks

Key Features:

• • Classification of tumor subtypes (PDM vs SCM)
• • Multiple ML approaches including SVM, XGBoost, and few-shot learning
• • Feature selection and dimensionality reduction techniques
• • 5-fold cross-validation evaluation

Results: Achieved 69.17% accuracy with Logistic Regression using Tumor Mutation Burden, copy number variations, and missense mutations with metaheuristic optimization for feature selection.