Engineered an AI-driven analytical and decision-support platform for investigating hydraulic behavior, predicting pressure dynamics, and evaluating operational strategies across complex water distribution systems.
Developed predictive machine learning workflows to model nonlinear relationships between network operating conditions and pressure behavior. Multiple regression strategies, including Random Forest and Decision Tree models, were evaluated through structured validation and performance analysis to investigate prediction reliability and model stability.
Pressure prediction experiments achieved an R² score of approximately 0.62. Residual behavior and performance metrics including MAE, MSE, and RMSE were analyzed to identify prediction limitations and operating conditions associated with reduced model reliability.
Designed a multi-scenario PRV optimization framework to investigate the trade-off between pressure reduction and low-pressure risk. Manual, randomized, and grid-based optimization strategies were comparatively evaluated, identifying operating configurations with average network pressure near 44.10 under controlled low-pressure penalty conditions.
Integrated hydraulic simulation, predictive modeling, optimization, and interactive visualization into a unified engineering analysis environment. Pressure trends, PRV behavior, residual patterns, and optimization outcomes were transformed into interpretable analytical insights for infrastructure performance assessment and operational decision support.
Technologies: Python • WNTR • EPANET • Scikit-learn • Random Forest • Decision Tree • TimeSeriesSplit • Predictive Modeling • Optimization • Plotly • Engineering Analytics
Engineered a real-time computer vision system for interpreting human movement through skeletal pose representation and temporal deep learning.
Designed a pose-based motion processing pipeline using 33 skeletal landmarks extracted through MediaPipe Pose. Body-centered normalization was applied to reduce positional variation and create more consistent motion representations across changing subject locations and camera perspectives.
Developed stacked LSTM architectures to analyze movement across 30-frame temporal sequences. This sequence-oriented approach enabled activity classification based on motion evolution rather than isolated visual frames, supporting recognition of walking, running, sitting, and standing behaviors.
Investigated temporal motion patterns and similarities between activity classes with overlapping skeletal configurations. Sequence segmentation, landmark preprocessing, and continuous buffering strategies were designed to maintain stable recurrent model inputs during live inference.
Integrated video acquisition, pose estimation, sequence construction, and neural network inference into a continuous real-time processing pipeline. An interactive monitoring environment was developed for prediction feedback, confidence inspection, skeletal visualization, and activity-state tracking.
The project demonstrates the use of temporal representation learning for transforming raw skeletal coordinates into interpretable motion sequences suitable for real-time behavioral analysis.
Technologies: Python • OpenCV • MediaPipe Pose • Deep Learning • LSTM • Recurrent Neural Networks • Pose Estimation • Temporal Sequence Modeling • Computer Vision • Real-Time AI Systems
Developed a scientific image compression and analysis platform for investigating the relationship between dimensional reduction, information preservation, and reconstruction quality in medical imagery using Singular Value Decomposition.
Implemented a low-rank approximation framework supporting medical DICOM images and standard image formats. Compression behavior was controlled through configurable rank selection and singular value energy-retention strategies.
Designed adaptive rank-selection mechanisms based on retained singular value energy, allowing compression configurations to respond to the information distribution of individual images rather than relying exclusively on fixed rank values.
Conducted singular value spectrum analysis to investigate information concentration and the contribution of dominant components to reconstructed image quality. PSNR and SSIM were integrated as quantitative evaluation metrics for analyzing reconstruction fidelity and structural preservation across compression configurations.
Investigated compression-quality trade-offs through comparative rank experiments and combined energy-retention analysis with reconstruction metrics to provide a broader assessment of information preservation.
Developed comparative visualization workflows for original and reconstructed images, intensity distributions, histograms, and singular value spectra. Batch-processing and automated evaluation capabilities were also implemented to support consistent analysis across multiple images and experimental configurations.
The project transformed SVD-based compression from a standalone mathematical technique into an interpretable experimental environment for studying low-rank representation, structural preservation, and image compression behavior.
Technologies: Python • Singular Value Decomposition • Low-Rank Approximation • DICOM • Medical Image Processing • PSNR • SSIM • NumPy • Scientific Computing • Image Quality Analysis
Developed an applied machine learning framework for investigating industrial process behavior, comparing predictive strategies, and supporting optimization-oriented operational analysis.
Designed structured analytical workflows combining preprocessing, feature analysis, predictive modeling, and comparative evaluation to investigate relationships between operational conditions and process behavior.
Developed and compared Logistic Regression, K-Nearest Neighbors, and Decision Tree models to analyze differences in predictive behavior, decision characteristics, and sensitivity to operational variation.
Rather than relying on a single predefined algorithm, model selection was approached as a comparative analytical problem. Performance behavior, prediction errors, and model stability were evaluated to investigate trade-offs between simplicity, interpretability, and predictive capability.
Applied hyperparameter optimization and structured validation strategies to reduce dependency on manually selected configurations and improve experimental consistency.
Model outputs were translated into comparative analytical insights designed to support evidence-based model selection and optimization-oriented decision-making.
The analytical pipeline was structured as a reusable experimental framework for iterative model comparison, performance evaluation, and industrial machine learning analysis.
Technologies: Python • Scikit-learn • Machine Learning • Logistic Regression • K-Nearest Neighbors • Decision Trees • Feature Engineering • Hyperparameter Optimization • Model Validation • Predictive Analytics
Developed and comparatively analyzed fundamental sorting and search algorithms with a focus on computational complexity, execution behavior, and performance variation across changing input conditions.
Implemented Merge Sort, Quick Sort, and Heap Sort in C# and investigated their runtime characteristics under different input sizes and data distributions.
Conducted time and space complexity analysis to examine theoretical performance characteristics and compare them with observed execution behavior.
Developed search implementations including Binary Search and hash-based retrieval approaches to investigate differences in lookup strategies and computational efficiency.
Designed comparative benchmarking experiments to evaluate execution behavior across algorithmic approaches and analyze how input scale and structure influence practical performance.
The project focused on connecting theoretical complexity analysis with empirical benchmarking, providing a structured understanding of algorithm selection and computational trade-offs.
Technologies: C# • Algorithms • Data Structures • Computational Complexity • Performance Analysis • Big-O Analysis • Benchmarking