GraphLingo defines a language for graphs allowing the creation and manipulation of graphs along with some commonly used graph algorithms. The entire project was created in Coq - The formal Proof Assistant, including defining the basic graph structure and the graph algorithms because Coq gives us the capability verify theorems and prove correctness. This helps extensively in creating robust algorithms, softfware, and compilers.
Created an integrated calendar application that presents a novel solution by harmonizing heart rate data, music listening habits, and personal scheduling within a unified interface. The idea is to investigate how past data can be levaraged to make better future decisions and the concept of "Quatified self". Contributed to the HCI community by creating a seamless user experience with ubiquitous design, providing a solution to the current fragmented system, and promoting the right use of data being tracked everyday. The calendar was created in ReactJS and the data was sourced from Apple watch and users' Spotify accounts. Gained experience with formative studies, Spotify API Interface, Apple Health Data Parsing, Data Analytics, and user studies.
The goal was to improve automatic license plate recognition by finding the right pipeline and understanding what helps the model specifically for low quality input images from pedestrian cameras. Tried to improve the accuracy of existing models with different combinations of OCR, super resolution, and preprocessing to see which one performed the best. Best Performing Model was: YOLOv7 for vehicle and LP object detection + Real-ESRGAN and OpenCV for image preprocessing to cropped LP images + Fine Tuned EasyOCR model for character recognition on the UFPR-ALPR dataset
Analyzed how debias are the Mixture of Expert models compared to Bert. The Debiasing techniques used were SentenceDebias and Iterative NullSpace Projection INLP. The datasets used were SEAT and GLUE dataset. Evaluation was done by the Sentence Encoder Association Test across gender, race, and religion. Switch transformer was first fine tuned on the GLUE dataset before adding any bias and measuring its prediction power. This project's goal was to evaluate models biasness and understand more what methods can be taken to make models less biased such that their prediction ability is not affected. Skills: Hugging Face, Google Switch Transformer, Mixture of Expert models
Suggested improvement on the existing Pandemic Simulator project by Dr. Peter Stone in "Reinforcement Learning for Optimization of COVID-19 Mitigation Policies" by suggesting to add the concept of job flexibility into the simulator. Modeled this parameter into the simulator and tested how the infection rates and hospitalization curves changes based on stay at home restrictions and analyzed using the models response, the benefits of having hybrid or work from home models. These type of simulator can be helpful for policy makers to weigh the pros and cons by seeing the impact before actually changing the policies.
Implemented a mirror server world online. Went through the entire process of design and implementation for this metaverse including ideation, user surveys, observations, prototyping, front-end, and back-end development. This is a multi-collaborative 3D art gallery environment hosted on a web server. x3d and gltf models have been used, and the back-end development has mostly been done in NodeJS and JavaScript. Here is the github link: The Art Metaverse
Check out my Wordpress blog for some HCI related writings. Samridhi Roshan's HCI Thoughts
This project is an online web pantry app that allows users to keep track of their pantry items and nutrition. There is also an easy and insightful graphical representation of pantry data for the users. MySQL, NodeJS, JavaScript, and ejs have been used. Here is the github link (does not have database information): FoodVentory
Implemented this personal webpage in Jekyll using HTML, CSS, and JavaScript. This project has been made accessible for people with any kind of color blindness including Protanopia, Deuteranopes, Tritanopes, and even Monochromacy commonly known as complete color blindness. This was the motivation behind creating the webpage in grey scale, using larger font family and sizes, using common icons and patterns, distinguishing sections clearly, and providing ease for navigation.
Implemented a spacecraft game in Unity. Used good software practices like encapsulation and code reusability. The concept of separation of concerns, object pooling, and flyweight pattern has also been leveraged. Contains enemy bullets and boss feature, a scoring system, level system, and it looks beautiful!
Implemented an external sorting algorithm for binary data in Java. Performed disk I/O on random access file and used Heap Sort algorithm for sorting.
Designed a working prototype for a hiking app. Used proper HCI principles including wireframes, storyboards, and the systematic cycle of design, analyze, prototype, and evaluate. Went through several iterations of ideation, design, and performed user evaluations as a group. Check out the final product here: Hike Advisor
Designed and Prototyped a simple events app for the city. Came up the problem statement and idea for this app. Check the prototype!
Implemented a simple customizable job control shell in C using the principles of process management and job control in a UNIX-like operating system.
Implemented a simple Mark-and-Sweep Garbage Collection algorithm in python.
Implemented an efficient Fork-Join Framework using the work stealing and thread safe approach.
Implemented an Assembler in C which supports a subset of MIPS32 languages (executed on Linux).
Programming lead - Semester Drone Project - Built a prototype of a drone where I used MATLAB to get meaningful graphs and to calculate the range for the drone.