Projects

• Assessed how well protein language models (ESM-3, ESM-2, ESM-C, ProstT5) capture family-level separation of targets, without explicit training.
• Dataset: Database of Useful Decoys – Enhanced (DUD-E).

• End-to-end computational pipeline classifying breast tumors by estrogen-receptor (ER) status from TCGA RNA expression data, for gene-signature discovery and drug/perturbagen matching against the iLINCS database.
• Applied gene filtering and selection using statistical testing with cross-validation, comparing nearest-centroid, logistic regression, lasso, and neural network models.
• Built an R Shiny application that computes per-gene differential-expression signatures from user-uploaded data, fetches concordant signatures through the iLINCS API, and renders a clustered heatmap.
• All methods classified ER status with >93% accuracy, with the neural network reaching 95.5%.

• Modeled molecules as graphs using Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) to classify different types of enzyme molecules.

• Semantic segmentation of retinal blood vessel images using the U-Net architecture.
• Trained different U-Net configurations (2, 3, and 4 blocks), reaching a highest dice score of 0.8.

• Fine-tuned pretrained vision models (ResNet-50, VGG16) to classify malaria cell images and identify infected cells.

• Embedded AI system for detecting bees, monitoring their behavior and health status, and suggesting possible solutions directly to farmers.
• Trained YOLOv5 for bee detection on a Raspberry Pi with Pi Camera, and fine-tuned ResNet-50 to classify disease from images.
• Built at Deltathon, Delta 3.0, Purwanchal Campus, IOE — received the Best Hardware Project Award.

Undergraduate research, advised by Prof. Bal Krishna Nyaupane

• Developed a sentence autocompletion tool for the Nepali language.
• Evaluated transformer and sequence-to-sequence models (LSTM, GRU) for sentence autocompletion in Nepali.
• Prepared a curated dataset of Nepali text for training NLP systems
• Trained and evaluated Word2Vec and transformer-based embeddings for modeling the Nepali language.

• A platform connecting patients in need of blood with donors who can post their availability.

• An application for investors to maintain a stock portfolio, receive profit and loss notifications, and discuss market queries in a discussion forum.