Nischal Bhattarai

I led rigorous testing efforts and played a pioneering role in the development of innovative AI/ML algorithms with the primary goal of significantly enhancing GPS positioning accuracy. I engineered state-of-the-art algorithms to solve intricate non-linear GPS navigation equations and improve triangulation techniques, pushing the boundaries of navigation technology. Additionally, I conducted extensive simulations of satellite flight paths and performed real-time data analysis, contributing to a deeper understanding of satellite positioning.

I spearheaded the creation and management of automation scripts and Singularity containers to optimize job automation and parallelism in HPC clusters. This involved installing, configuring, and maintaining scientific software with MPI support using manual compilation and tools like EasyBuild. Additionally, I developed Python packages for automating documentation, TFLOPS estimation, data visualization, and cluster analysis in SLURM clusters. Notably, I built and maintained a WordPress-based documentation website hosting 100+ HPC software packages, integrating SLURM scripts and enhancing user accessibility to crucial HPC resources.

The Navigator is a python based GNSS library and toolkit tailored for GNSS PVT solutions. It provides a uniform object-oriented API to do basic GNSS data processing, including RINEX parsing, satellite position calculation, and user PVT calculation. The library is well documented and easy to use. It also provides a set of CLI tools for RINEX data collection from public FTP servers.

The goal of this project was to determine the fractional mass of cold dark matter with subhalo substructure using a simulated gravitational lensing dataset. The model of choice was a Convolutional Neural Network (CNN), which contains 21 CNN operators distributed over 7 residually connected layers. Each layer has 3 CNN sub-layers. In each block, the 3 sub-layers are further residually connected to each other using a scaler gate, which provides dedicated LR (Learning Rate) as well as allowing gradients to flow in each sublayer. The convolution layer outputs are fed into 6 linear 3-2 residually connected layers of the same structure. The model was trained for 10 epochs; however, the convergence was amazingly fast, achieved in just 2 epochs.

SlurmDocs is a Python package and CLI tool designed to streamline the process of documenting Slurm clusters. Whether you're managing a small research cluster or a large-scale HPC system, SlurmDocs automates the documentation process, saving you time and ensuring your cluster information is always up-to-date.