Suramya Pokharel

Puzzle Flow

This project focuses on solving jigsaw-like puzzle with identical square pieces all while enhancing the quality of the image using Flow Based Generative AI models and Vision transformer. The project can be further extended to restore shattered documents, DNA sequence assembling in genomics or in room layout arrangement in real estate domain.

An Electronic Watch Dog

This project consisted of an Arduino powered robotic vehicle with an attached wireless camera that could send live video to server where faces would be detected using Haar cascades and recognized using FaceNet model. The application of this project can be seen in sectors of home security and surveillance.

Hospital Management System

This project revolved around usage of database system to store information of various faculties of a hospital like doctors, patients, accounts and so on. Website for the project was using HTML, CSS and Django while SQL was used as the database system.

Principal Component Analysis

This project focuses on finding out the main attributes used to seprate class labels from a dataset. Principal components of the "IRIS", "Haberman's Survival" and randomly sampled dataset were found out.

Naive Bayes and its Variants

This project utilized variants of naive bayes classifier like simple, categorical and hybrid to classify stars into their appropriate types using "STAR" dataset.

K-means to find proportions

K means classifier was used to calculate the proportion of marble in a dusty image and the area of lake in an geographical birdvew image. Various relevant data about the images were gathered in the process.

KNN to classify stellar bodies

This project used KNN to classify the labels of Stellar Classification Dataset - SDSS17. Euclidian, Manhattan and Chebyshev distances were used to find the most optimal classifier. Cross validation was done to achieve high accuracy.

Building a brain - ANN

An Artificial Neural Network was built to classify the labels of MNIST dataset. Various changes were brought to the ANN like usage of regularizarion, addition of extra layer, changes of total neurons to evaluate accuracy and the results were documented.

Decision Tress to seperate glasses.

Decision tree classifiers with different splitting criteria varying from GINI, Entropy to Information gain were used on the "Glass Identification" dataset

Dimensional Labyrinth - KD tree

This was a case study done on KD tree. It involved realizing a KD tree from scratch and evaluating various properties of the tree through ablation studies on various datasets. The process was thoroughly documented.