Key Responsibilities:

• Investigating cross-session classification of left and right-hand motor imagery from EEG, with emphasis on participant variability and model interpretability.
• Developed leakage-controlled CSP-LDA and EEGNet baselines on BCI Competition IV Dataset 2a; a three-participant pilot yielded mean balanced accuracies of 78.24% and 75.23%, respectively.
• Currently designing a channel frequency attention model to identify informative EEG channels and frequency bands.

Skills Applied: EEG Signal Processing, Brain–Computer Interfaces, CSP-LDA, EEGNet, Cross-Session Evaluation, Model Interpretability

Key Responsibilities:

• Investigated whether a Vision Transformer paired with an efficient recurrent model can detect and classify real-world violent anomalies in long, sparsely labeled surveillance video.
• Lead contributor to model training, comparative experiments, result visualization, and ablation studies. Designed and implemented the complete end-to-end pipeline, including training and evaluation notebooks, temporal segmentation and augmentation, ViT-small feature extraction, and attention-enhanced SRU++ modeling.
• Achieved 95.14%, 92.88%, and 93.73% test accuracy on UCF-Crime, RWF-2000, and SCVD, respectively, with the highest reported AUC among the compared methods on all three datasets.

Skills Applied: Deep Learning, Computer Vision, Video Anomaly Detection, Vision Transformers, Attention Mechanisms, SRU++, Video Segmentation and Augmentation, Multiclass Classification, Feature Clustering, Model Training, Ablation Studies, Performance Evaluation

Key Responsibilities:

• Investigated whether a computationally lighter, non-neural framework could accurately predict snoRNA–disease associations without relying on graph neural networks.
• Co-developed a two-stage GBDT-to-SVM framework integrating snoRNA functional similarity, disease semantic similarity, and Gaussian interaction-profile kernels; used K-means-based proportional negative sampling to address class imbalance, contributed to methodology and analysis, and independently verified top-ranked predictions using RNADisease v4.0 and PubMed across 12 diseases.
• Achieved mean AUROC 0.96, AUPRC 0.95, accuracy 0.95, and F1-score 0.94 on MDRF under five-fold cross-validation, with higher reported AUROC/AUPRC than seven comparison methods; also evaluated generalization on LSGT and PsnoD.

Skills Applied: Bioinformatics, Feature Engineering, Ensemble Machine Learning

Key Responsibilities:

• Investigated which language and CPU architecture minimizes the energy and carbon footprint of statistical and ML workloads.
• Designed a cross-platform energy-measurement framework using OS-native tools (powermetrics, powerstat, and HWiNFO), isolating task energy via idle-power subtraction; ran 72 benchmarks per OS across Windows, Linux, and macOS.
• Tests show ARM drew approximately 100× less idle power than x64 (approximately 0.05 W vs. 4–9 W), with energy scaling by dataset size; Python was most efficient on ARM and R on x64.

Skills Applied: Green Computing, Energy Profiling, Cross-Platform Benchmarking, Python and R, Hardware-Aware Performance Analysis

Key Responsibilities:

• Investigated whether coupling K-means clustering with Random Forest could improve classification performance relative to a conventional Random Forest baseline on UCI benchmark datasets.
• Lead contributor of methodology, implementation, and evaluation. Designed and implemented the RFWOC pipeline, using Elbow-method selection of k, cluster-specific Random Forest classifiers, and comparative evaluation; led experiment execution and result visualization, authored the Methodology section, and co-authored the Results section.
• Outperformed the classical Random Forest baseline on most evaluated benchmarks, including 88.6% accuracy on Heart Disease, 76.9% on Diabetes, and 95.2% F1-score on Breast Cancer.

Skills Applied: Machine Learning, Ensemble Learning, K-Means Clustering, Random Forest, Data Preprocessing, Model Evaluation, Experimental Analysis

Key Responsibilities:

• Investigated whether a language model could flag weak passwords by modeling character-sequence predictability.
• Developed a GPT-2-based proof of concept using a corpus of 1.92M unique passwords prepared from 10M+ RockYou and SecLists entries; implemented character 1–3-gram tokenization and perplexity-based scoring.
• In exploratory tests, perplexity separated trivial passwords (approximately 2–5) from complex ones (greater than 50); documented dataset-scale and model-capacity limitations.

Skills Applied: Password Security Analysis, Deep Learning, Transformer Language Models (GPT-2), Character N-gram Tokenization, Perplexity-Based Evaluation