Projects

Neural Network from Scratch (MNIST)

Fully connected neural network built using only NumPy, no PyTorch or TensorFlow. Every component, forward pass, backpropagation, and gradient descent, implemented manually. 96.57% test accuracy on 10,000 MNIST test images.

Brain MRI Tumor Classification

ResNet-50 transfer learning pipeline classifying brain MRI scans into four categories (Glioma, Meningioma, Pituitary, No Tumor) using two-phase training and noise-robust augmentation. 87.71% accuracy, ROC-AUC 0.9711, with Grad-CAM interpretability and a Gradio interface at 0.84ms per image inference.

Sepsis Early Warning System

ICU decision support tool predicting sepsis onset from over 80 engineered temporal features on a 43:1 imbalanced dataset. 5-layer DNN with weighted sampling and G-mean threshold optimization, 97.38% accuracy, ROC-AUC 0.9238, 0.016ms per sample inference.

Decision support tool only. Not a substitute for clinical judgment.

Monte Carlo Neutron Transport Simulation

Stochastic neutron transport simulation in C++ modeling scattering, absorption, and fission to estimate a reactor's multiplication factor (k-eff). Runs 10,000 independent simulations with generation-based population control, achieving k-eff convergence within 0.1%. Results analyzed and visualized in Python.