Adaptive Transformer Architectures
Deep Learning · Transformers ·
Time-Series Modeling · Model Optimization

The Challenge
Conventional Transformer architectures face computational and modeling limitations when processing complex, long-range temporal sequences.

My Work
Developed and investigated novel Transformer architectures for sign language machine translation:
- ADAT: An adaptive Transformer combining convolutional feature extraction, LogSparse attention, and adaptive gating.
- GLoT: A gated-logarithmic Transformer designed to model long-range temporal dependencies.
- HATL: An adaptive transfer learning framework that progressively unfreezes pretrained layers based on training behavior.

Research Results
ADAT achieved an average 21% reduction in training time across three sign-to-gloss-to-text benchmarks while improving translation quality against the evaluated baselines.

Computer Vision & Sign Language Recognition
Computer Vision · YOLO ·
Transfer Learning · Multimodal Data

The Challenge
Accurate sign language recognition requires robust visual feature extraction and representative training data, particularly in specialized and low-resource domains.

My Work
- Contributed to VisioSLR, a computer vision framework evaluating fine-tuned YOLO architectures for sign language recognition.
- Developed MedASL, a medical-domain American Sign Language dataset containing 700 annotated sign videos, supporting research in healthcare-related sign language translation.

Research Results
The VisioSLR evaluation reported 97.4% mean average precision and 11 ms recognition time using YOLOv8m. MedASL was introduced and evaluated as a benchmark in subsequent sign language translation research.

Privacy-Preserving & Distributed AI Systems
Edge AI · Cloud Computing · Blockchain · AI Governance · MLOps

The Challenge
AI systems handling sensitive user data must address privacy, governance, computational efficiency, and traceability alongside model performance.

My Work
Developed a consent-aware, end-to-end AI framework integrating edge computing, cloud-based model development, and blockchain-supported auditability.
The architecture incorporates data preprocessing, model inference, consent management, model lifecycle operations, and traceable system events.

Research Results
Evaluated the framework using Transformer-based translation models, measuring translation quality, training efficiency, communication time, and end-to-end performance in an experimental environment.

NLP-Based Emotion Analytics
NLP · Machine Learning · Sentiment Analysis ·
Data Annotation

The Challenge
Understanding public emotional responses to major events requires extracting meaningful patterns from large volumes of unstructured social media data.

My Work
Contributed to ML-NLPEmot, a machine learning and natural language processing framework for event-based emotion detection.
The research involved social media data collection, emotion annotation, feature extraction, and comparative evaluation of machine learning models using Plutchik's emotion categories.

Research Results
The published study reported emotion prediction accuracy of up to 95.5% in its evaluated setting. The associated annotated dataset is available publicly for research use.