As a member of the analytics team at eGain, I help maintain and modernize key backend systems that support customer insights and reporting. My work involves maintaining legacy codebases, resolving system bugs such as login pipeline issues, and migrating infrastructure to more stable and modern architectures. I help update existing code to support newer libraries, APIs, and platform standards, improving long-term reliability and performance. This role requires close collaboration and a strong focus on clean, maintainable code and system resilience.

As part of my Master’s project at UC Santa Cruz, under the guidance of Professor Roberto Manduchi, I developed an iOS application that improves pedestrian localization accuracy in dense urban environments where GPS often fails.

The system combines GPS, RoNIN (a neural inertial navigation algorithm), and particle filtering into a unified positioning framework. I expanded the codebase by integrating detailed environmental constraints, including hand-labeled maps of buildings, sidewalks, and crosswalks. I also built a real-time simulation engine, designed a ground truth generation pipeline, and led field data collection across downtown San Francisco.

The fusion model reduced mean positioning error by more than 50 percent compared to GPS alone and increased sidewalk identification accuracy from 47.5 percent to 80.5 percent. It consistently outperformed GPS and maintained consistent accuracy even when the phone was carried in a pocket. Unlike standalone methods, our approach was able to determine not only the user's location but also which side of the street they were on.

This project delivers a practical and highly accurate solution for pedestrian navigation in urban environments, with strong potential to support accessibility tools for the visually impaired and improve general-purpose mobile positioning.

As a Teaching Assistant at the University of California, Santa Cruz (UCSC), I am currently engaged with the CSE 102 course, Introduction to Analysis of Algorithms. In this role, I have the opportunity to guide and support approximately 70 students, a responsibility that I undertake with great enthusiasm and dedication.

My key responsibilities encompass a broad range of educational tasks, including grading assignments, conducting weekly office hours, and leading instructional sections for two hours each week. These tasks not only allow me to interact closely with students but also provide a platform for facilitating their learning and understanding of complex algorithmic concepts.

While I already possessed a strong foundation in algorithms, this role has significantly deepened my expertise in the subject. Teaching these concepts to students has enabled me to refine my understanding of the topic, making me more adept at tackling intricate problems and explaining them in an accessible manner.

Additionally, this position has been instrumental in enhancing my communication and leadership skills. Regular interactions with students and the responsibility of leading educational sections have honed my ability to convey complex information clearly and effectively. This role has also nurtured my skills in leadership, as I guide students through their academic journey in this challenging subject.

Through this role, I am not only contributing to the academic growth of students at UCSC but also continuously developing my own skills in algorithm analysis, communication, and leadership.

In collaboration with two professors, Martha Zuniga and Sangwon Hyun, at the University of California, Santa Cruz (UCSC), I worked on an innovative project to develop an adaptive, automated flow cytometry gating software. This cutting-edge tool leverages Gaussian Mixture Modeling to accurately identify and cluster cell populations, a critical step in flow cytometry analysis.

Our approach allows for more precise and efficient identification of cell subsets. This methodology not only enhances the quality of the data obtained but also substantially improves its replicability.

My responsibilities primarily included writing the code, brainstorming ideas, and presenting progress of the project every week.

I presented our findings in a conference during the winter of 2024.

Here is a link to some of what we have done so far (without any code)

Note that you have to download the html file to view it.

https://drive.google.com/file/d/1bxDplpAyHMo_DTr4_PWPvA1qY0AE3OOf/view?usp=sharing

During my internship at Wisepal, I focused on optimizing machine learning algorithms and enhancing the company's large language model capabilities. I successfully optimized a central algorithm, reducing its runtime by 75% - from 3-4 hours to approximately 1 hour through parallelization.

I improved the accuracy of the companies large language model from 97% to 99%, representing a 66% reduction in the remaining error margin, significantly enhancing the model’s performance. During training I noticed erroneous data points. Through my own volition I wrote a script to fix/remove errors in the underlying data. These validation checks improved/removed over 14% of the dataset.

A key aspect of my role involved researching, training, and optimizing Large Language Models (LLMs) for email information extraction. These models included FLAN-T5, BERT, Mistral 7B, T5, and T5v1.1. In this process, I implemented advanced techniques including LoRA (Low-Rank Adaptation) and prefix tuning to enhance model performance.

I conducted comprehensive comparative research on various model architectures to determine the optimal solution for the company's specific needs. This research involved exploring and implementing multiple fine-tuning approaches, including traditional fine-tuning, prompt tuning, prefix tuning, LoRA / QLoRA (Quantized Low-Rank Adaptation), and sparse fine-tuning.

Throughout my internship, I regularly presented cutting-edge research findings to the team and successfully integrated these insights into the company's codebase. My contributions helped advance Wisepal's machine learning capabilities. This experience not only enhanced my technical skills but also improved my ability to communicate complex technical concepts effectively.