I am a Machine Learning Engineer focused on developing robust Computer Vision and NLP systems. My work emphasizes the transition of deep learning models from experimental research to stable, production-ready inference services.

I am currently advancing my technical expertise at Islamic Azad University, Science and Research Branch.

• Computer Vision & XAI: Custom Native PyTorch CAM (Class Activation Mapping), Vision Transformers (ViT), CNN Architectures (EfficientNet, ResNet), CLAHE Preprocessing.
• Natural Language Processing: Transformer Architectures (T5, RoBERTa, DistilBERT), Sequence Modeling, Sentiment Analysis, Grammatical Error Correction.
• Machine Learning & Stats: XGBoost, LightGBM, Feature Engineering, Advanced EDA, Regression & Clustering.
• Operations & Engineering: Docker, Git/GitHub, Linux/Bash Scripting, SQL (Window Functions), Gradio & Streamlit.

• Engineered a multi-label classification system for 14 pathological conditions using EfficientNet-B0, achieving a mean AUC-ROC of 0.825.
• Developed a Custom Native PyTorch CAM implementation using forward and backward hooks to provide clinical interpretability, bypassing dependency conflicts inherent in standard 3rd-party libraries.
• Integrated CLAHE contrast enhancement for medical image normalization and resolved ASGI/Uvicorn event loop conflicts for stable cloud deployment.

• Implemented a Vision Transformer (ViT) using PyTorch for 101-class image classification.
• Optimized the model using advanced data augmentation and transfer learning on a dataset of 100k+ images to ensure high generalization.
• Achieved 95% training accuracy and 90% test accuracy, deployed via a scalable web interface.

• Fine-tuned a T5-Small Transformer on a 500MB Lang8 dataset for end-to-end linguistic error correction.
• Achieved a 67.42 SacreBLEU score and a 0.857 ROUGE-1 score, demonstrating high accuracy in sequence-to-sequence generation.
• Publicly hosted with interactive Gradio and Streamlit interfaces for real-time inference.

• Machine Learning Intern | BecomeExpert: Engineered deep learning pipelines for structural crack detection (95%+ accuracy) and developed high-precision pricing models using XGBoost and LightGBM.
• Data Analyst | Ravan Rail: Managed operational data for 200+ employees, producing structured financial reports and auditing payment records.

• LinkedIn: erphan-rajai
• Hugging Face: itserphan
• Email: itserphan@gmail.com
• GitHub: ErphanRajai