Computer Science undergraduate at the University of Tehran (GPA 17.56/20). I'm interested in deep learning for medical and biological data, especially imbalanced datasets, and in writing careful, tested, reproducible ML code.
- numpy-autograd: a tensor-level autograd engine and neural-network library in pure NumPy (Linear, Dropout, SGD / Momentum / Adam). Verified against PyTorch with 75 tests and CI; an MLP trained with it reaches 89.0% test accuracy on Fashion-MNIST (PyTorch: 88.7%).
- DermaMNIST transfer study: does ImageNet pretraining help on small, imbalanced skin-lesion images? A custom CNN vs. ResNet-18 from scratch vs. pretrained ResNet-18, with the training loop, Focal Loss, Grad-CAM and temperature scaling written by hand in PyTorch (37 tests, CI). The pretrained model reached test Macro-F1 0.692 (vs. 0.520 from scratch, 3 seeds); melanoma recall stayed at 0.55, and the official split shares lesions between train and test, documented as a limitation.
- Traffic Severity Prediction: 4-class severity prediction on 100K imbalanced samples. Leakage-safe pipeline, SMOTE, feature selection (35→25) and SHAP. XGBoost reached Macro F1 0.577 (vs. 0.478 for the linear baseline).
- x86-64 To-Do Tracker: command-line to-do app in NASM assembly using raw Linux system calls.
Artificial Intelligence · Machine Learning · Theory of Computation
- ML: PyTorch · NumPy · scikit-learn · XGBoost · SHAP · pandas · matplotlib
- Engineering: Python · C/C++ · Git · GitHub Actions · pytest · Linux
