Skip to content
Ubaid Ullah

Academic Work

Research

Thesis work, research interests, and ongoing projects at the intersection of AI and intelligent systems.

MS Thesis

Student Information System

MCS · Abdul Wali Khan University, Mardan

Focus Areas

Research Interests

Artificial Intelligence

Designing intelligent systems that reason, learn, and adapt to real-world problems.

Machine Learning

Applying statistical learning methods to extract actionable insight from structured and unstructured data.

Deep Learning

Building and evaluating neural network architectures for prediction, classification, and assessment tasks.

Intelligent Systems

Engineering systems that combine AI models with robust, production-grade software architecture.

Data Science

Turning raw data into rigorous, reproducible analysis that informs decisions.

Selected Work

Research Projects

Credit Risk Prediction using Deep Learning

Featured

Problem Statement

Traditional credit scoring models struggle to capture complex, non-linear relationships in borrower data, leading to suboptimal default prediction and missed risk signals.

Methodology

Designed and trained deep neural network architectures (including feed-forward and regularized deep nets) on real-world credit datasets, benchmarking against logistic regression, random forests, and gradient boosting baselines. Applied feature engineering, class-imbalance handling, and k-fold cross-validation to ensure robust generalization.

Results

The deep neural network model outperformed traditional baselines on key metrics (AUC-ROC, F1-score), demonstrating the viability of deep learning for credit risk assessment. This work formed the basis of the MS thesis at SZABIST University.

Technologies Used

Python TensorFlow Scikit-Learn Pandas NumPy

DASS-42 Mental Health Assessment using Machine Learning

Featured

Problem Statement

Manual scoring and interpretation of the DASS-42 (Depression, Anxiety and Stress Scale) questionnaire is time-consuming and can benefit from automated, data-driven severity classification.

Methodology

Built supervised machine learning classifiers trained on DASS-42 response data to automatically predict severity levels across depression, anxiety, and stress sub-scales, with exploratory analysis of feature importance across questionnaire items.

Results

Demonstrated that machine learning classifiers can reliably automate DASS-42 severity scoring, offering a scalable tool for mental health screening support.

Technologies Used

Python Scikit-Learn Pandas Matplotlib

Bone Classification using Deep Learning

Featured

Problem Statement

Manual classification of bone types and conditions from medical images is labor-intensive and subject to inter-observer variability.

Methodology

Developed convolutional neural network models for image-based bone classification, applying data augmentation and transfer learning to improve performance on a limited medical imaging dataset.

Results

Achieved strong classification accuracy using transfer learning, highlighting the potential of deep learning to support radiological workflows.

Technologies Used

Python TensorFlow Keras OpenCV