Rishit Kar

Deep learning framework for robust and explainable multi-omics cancer subtype classification

College Innovative Product Development Project (Supervised by Prof Ruhina Karani) Designed and developed OMNIGATE, a novel deep learning framework for multi-modal cancer subtype classification using integrated omics data (mRNA, miRNA, CNV, Methylation). Introduced a dynamic context gating mechanism that adaptively weights modality importance per sample, improving robustness over traditional feature concatenation methods.

Implemented a multi-objective loss function combining Focal Loss (handling class imbalance), Orthogonality & Alignment Loss (ensuring meaningful latent representations), and Sparsity & Entropy Regularization (encouraging decisive modality selection).

Built an explainable AI pipeline to extract Top-K biomarkers using gradient-based sensitivity analysis and visualize modality importance, enabling interpretable and clinically relevant predictions.

Developed using PyTorch and scikit-learn for scalable experimentation on multi-omics datasets.

Developer-focused CLI/TUI clipboard manager with real-world adoption on PyPI

Developed Clipper-dev, a cross-platform clipboard manager enabling persistent history tracking, structured content organization, and efficient retrieval through commands like add, search, restore, and export across macOS, Linux, and Windows.

Designed an interactive Terminal User Interface (TUI) with features including real-time fuzzy search, history browser, and a statistics dashboard to enhance developer productivity.

Published as a Python package on PyPI with 2300+ downloads, demonstrating practical adoption and usability in real-world developer workflows.

Implemented using Python with pytest-based testing and robust packaging for seamless distribution.

Deep learning ensemble for medical image classification using CNN, ResNet, and Graph Neural Networks

Built a comprehensive deep learning solution for pneumonia classification using chest X-ray images. Implemented custom CNN architecture alongside ResNet and Graph Attention Network (GAT) models for comparative analysis. Developed multimodal learning pipeline integrating spatial and attention-based features. Achieved high test accuracy with strong F1 scores for clinical reliability using TensorFlow and Keras.

High-performance reliable transport protocol with sliding window flow control

Engineered a reliable transport protocol from scratch using Python UDP sockets, incorporating TCP-like features including sliding window flow control and automatic retransmission. Implemented SHA-256 checksums for data integrity and concurrent packet processing. Achieved significant performance improvements over traditional stop-and-wait protocols through optimized throughput testing.

Privacy-focused VS Code extension for cleaning tracking parameters from URLs

Created and published a VS Code extension that automatically removes tracking parameters from URLs on paste, enhancing user privacy and link readability. Implemented offline functionality using regex patterns without external dependencies. Built using JavaScript and VS Code API, with comprehensive documentation and marketplace compatibility.