Personal Projects
- Graph-NERD
- NuQuant
- Keshik
- Super-Resolution GAN for Precipitation Downscaling
- Temporal Causal Graph Discovery in HPC Network Traffic Simulations
A Graph Neural Network (GNN)-based framework for relational causal discovery. It integrates FastGAE-based structural pruning to reduce the causal search space and improve the scalability of structural equation modeling on relational datasets.
A calibration-free, neuron-level quantization method for large language models based on Mixture-of-Experts (MoE) architectures. Evaluated on Mixtral 8x7B, NuQuant reduces model footprint for efficient LLM inference while maintaining strong downstream task performance.
A Python package for addressing class imbalance in tabular datasets using Denoising Diffusion Probabilistic Models (DDPMs). Built with PyTorch, it provides generative oversampling for applications such as fraud detection, anomaly detection, and customer churn prediction.
A super-resolution GAN for converting low-resolution 16x16 South Asian precipitation fields into 64x64 high-resolution data. The project improves the spatial fidelity of meteorological data for downstream forecasting and analysis.
A causal discovery project on multivariate temporal network traffic simulations from Argonne National Laboratory. Applied NAVAR and CausalFormer with Regression Relevance Propagation (RRP) to identify nonlinear causal relationships and construct temporal causal graphs for improved surrogate-model forecasting.