Pial Abedin
ML/AI Engineer
North South University
Project
Built an end-to-end pipeline on 5,351 patient records using pre-operative features only, with strict leakage prevention and SMOTE to handle 2.7% class imbalance. 97.5% accuracy, AUC = 0.813 — used SHAP for clinical explainability; deployed the final model
Project
Compared LSTM, Bi-LSTM with custom Attention, and Transformer architectures for stock price prediction; deployed as a live dashboard with real-time Yahoo Finance data. Bi-LSTM + Attention achieved best performance: RMSE = 4.13, R² = 0.9556 on out-of-sampl
Project
14-disease chest X-ray detector: trained DenseNet121 on 89,696 NIH ChestX-ray14 images, reaching mean AUC 0.7765; Grad-CAM heatmaps visualize model focus regions. Deployment: LLaMA (via Groq API) auto-generates structured radiology reports; full-stack dep