At IIIT Bhagalpur, I built a deep learning model to predict magnetic-material B-H loops using Princeton's MagNet database. I compared ANN, RNN, LSTM, and GRU models on a material they hadn't seen during training; the GRU achieved 59% lower MSE than the ANN.
I also built a four-agent research tool with LangGraph, FastAPI, and Streamlit, where agents retain state between steps and their status appears live. My projects include writing a feedforward neural network with backpropagation and gradient descent in NumPy, without ML libraries.
