Projects
Selected work across efficient language models, Transformer internals, higher-order learning, software security, computer vision, and full-stack engineering.
My public code and research repositories are hosted on GitHub · donnowhattodo ↗.
Reimplements the original encoder–decoder Transformer end to end—from token vocabularies, positional encoding, masks, and scaled dot-product attention through multi-head attention, residual normalization, label smoothing, Noam scheduling, and beam search.
Open repository →A memory-efficient Qwen2.5-7B-Instruct fine-tuning framework using 4-bit loading, QLoRA, gradient checkpointing, adaptive rank scheduling, SVD warm-start, gradient-noise-gated control, and reproducible GPU isolation.
Ongoing researchA leakage-controlled benchmark across nine interaction datasets comparing pairwise graph projection with incidence-based penalized hypergraph spectral clustering under matched-k evaluation.
View preprint →A CodeBERT-based pipeline for classifying secure and vulnerable source-code snippets using DiverseVul, with knowledge distillation to lower inference cost for practical code-scanning workflows.
Ongoing researchA Django social platform prototype for North South University students, with authentication, content sharing, topic-based help, real-time comments, and direct messaging. The prototype was not publicly released.
Open repository →Studies how random pixel blocking at 5%, 10%, and 15% affects CNN performance on Intel Image Classification, with training curves, test accuracy, loss analysis, and confusion matrices under regularization.
Open repository →An end-to-end multiclass NLP pipeline using 60,000 Stack Overflow questions to distinguish high-quality, community-edited low-quality, and closed low-quality questions.
Open repository →Fine-tuned ViT-B/16 for binary fracture detection on the MURA v1.1 radiographic dataset and evaluated model generalization using five-fold cross-validation.
Open repository →Compared linear and nonlinear SVM kernels across multiple text representations, reporting accuracy and F1-score alongside computational cost to study efficient alternatives to Transformers.
Open repository →Explored teacher–student training for sentiment analysis, MRPC, and CoLA, with reusable fine-tuning, distillation, and evaluation components.
Open repository →Used PCA and a residual U-Net with spectral normalization to reconstruct 31 spectral bands, evaluated with PSNR, SSIM, and Spectral Angle Mapper.
Open repository →Implemented image and video inference using pretrained YOLOv3 and YOLOv4 models on the COCO label space for real-time detection across 80 object categories.
Open repository →