Yash Vardhan Pansari's Projects
2D models for the design of a 20T Accelerator Magnet
Built this as a mother's day gift for my mom, but can mostly be used to make a painting-like animation for mostly any photo
Used Sparse Temporal Sampling on Georgia Tech's Let's Dance database and experimented with data augmentation before feeding it into a customised 3D Mixed Convolution ResNet 50 Model, trained to classify the randomly sampled, mute, video frames into 10 dance categories. Contested Georgia Tech's claim that the videos in their database were visually overlapping and thus indistinguishable based on frames.
Fall '21 NMEP Central Repo
This repo contains the completed projects and homeworks from my time in ML@B's NMEP (a student-run fundamentals of ML course to prepare for work on advanced industry and research projects in ML). This is meant for recruiters and employers, not for later students to plagiarize.
Git-inspired version control system designed and elaborately constructed in Java. This code was written in response to a project assignment in the DSA class at Berkeley but is published here for recruiters and employers, not future students.
A package to identify, separate, extract, and analyse strands, filaments, and cracks, from stitched microscopic cross-sectional images of Nb3Sn Magnet Superconducting Coils. This algorithm was developed by me in a research project for the US Department of Energy, in coordination with research scientists and engineers at CERN and Lawrence Berkeley National Laboratory's ATAP division.
Numpy like linear algebra library written in C, providing a 210x speedup. This code was written in response to a project assignment in the computer architecture class at Berkeley but is published here for recruiters and employers, not future students.
Built a Feature Pyramid Network with a ResNet-50 backbone with a custom Torch loss function for rotated bounding boxes. Used it for object detection and segmentation on adversarial noisy binary images of stars, achieving over 96% IOU accuracy.
Used self-supervised learning in tandem with various modern ResNet models; Trained them on common image databases such as CIFAR-10, MNIST, etc. using PyTorch and git; Built a RotNet model to extract features from diverse types of unlabelled images (without human supervision).
Used simulated annealing via Metropolis Monte Carlo algorithm to determine (unsupervised) clusters for wines on chemical constituency.
An autocorrected speedtyping game inspired by Typeracer.
Built low compute Spatio-Temporal Encoder-Decoder using time-series satellite imagery to improve weather now-casting for solar grid scheduling. Reached top 5 internationally in the ClimateHack.AI qualifiers.
Config files for my GitHub profile.