Humera Mirza

I've independently built and deployed end-to-end machine learning projects spanning computer vision, NLP, and recommender systems.

For radio galaxy morphology classification, I trained group-equivariant CNNs on 2,158 FIRST survey images to classify FRI, FRII, Compact, and Bent galaxies. My best model achieved 81% test accuracy and a Macro F1 of 0.81, and I used Grad-CAM and confusion-matrix analysis to interpret predictions.

I also built a content-based recommendation system for 1,500+ translated Chinese web novels. I created the pipeline from web scraping and cleaning through PostgreSQL storage, semantic embeddings, and a hybrid engine combining trope, description, and collaborative-filtering signals.

I work with Python, PyTorch, FastAPI, PostgreSQL, Docker, Gradio, and Hugging Face Spaces to take models from training and evaluation through accessible deployment.