During my time as a Research Intern on the SAINTES team at Microsoft Research (MSR), I implemented machine learning solutions to streamline release engineering workflows. I collaborated with the IDNA product team to explore AI-driven productivity enhancements for engineers and conducted interviews with on-call engineers to identify inefficiencies. I developed a novel model using Large Language Models (LLMs) to automatically label differences between test and production environments as "noise" or "critical," improving efficiency.

I fine-tuned GPT-3.5 and GPT-4 using Azure OpenAI Studio, achieving 98% accuracy in offline evaluations and consistently over 90% accuracy in live on-call sessions. This reduced labeling time from 10 hours to just 15 minutes, significantly enhancing developer productivity. I presented the results to the Corporate Vice President (CVP) and senior leadership in IDNA, and completed a successful tech transfer, where I transferred my model into the product team’s production workflow.

I am working on a project to enhance cross-domain recommendation capabilities using state-of-the-art language models. By introducing zero-shot and few-shot prompting techniques, I optimized the recommendation process across various domains. The project involves fine-tuning open-source models like LlaMA-2 with parameter sizes ranging from 7B to 70B. This process was conducted using parameter-efficient fine-tuning (PEFT) methods. A critical component of the project was the use of QLoRA for fine-tuning LLMs on dual-domain data, improving the models' recommendation performance. The fine-tuning process was powered by multiple A100 GPUs, ensuring efficient handling of large-scale data and model parameters. I utilized Weights and Biases for visualizing the fine-tuning process of QLoRA, allowing for a detailed and transparent view of the model's performance enhancements. This comprehensive approach demonstrated the potential of LLMs in revolutionizing the field of recommender systems through advanced transfer learning techniques.

I conducted reasearch that integrated the Rips Complex from computational geometry with deep learning techniques. The primary goal was to effectively transfer user preferences across varying domains, leading to the creation of novel user profiles tailored for personalized recommendations. The approach was particularly groundbreaking in addressing extreme cold-start scenarios, registering a substantial performance boost of over 20% compared to existing leading methods. This project involved meticulous fine-tuning using 5-fold cross-validation to enhance the model's efficacy on novel data. Technologically, it leaned heavily on Python and Tensorflow, with all complex computations being adeptly handled by the powerful Nvidia A100 GPU.

I leveraged Python, GloVe pre-trained embeddings, and BERT techniques from NLP to develop sophisticated domain similarity metrics. This led to the formulation of baseline cross-domain recommendation algorithms, taking advantage of the open-source Recbole-CDR library. The algorithms were rigorously assessed across a diverse array of 18 domain combinations, benchmarked against three industry-leading cross-domain algorithms. Intriguingly, the findings revealed that the recommendation efficacy was largely invariant to the specific domain combinations, a conclusion drawn using a paired t-test. For optimum efficiency, all computational processes and experiments were powered by the high-performance Nvidia A100 GPU.

I utilized Python, TensorFlow, and the Nvidia A100 GPU to develop a cross-domain recommender system using a conditional GAN. The project centered around the synthetic generation of target domain item embeddings, which were seamlessly infused with source domain data. Furthermore, these synthesized embeddings were combined with pre-existing user embeddings, providing enhanced personalized recommendations. As a testament to the robustness of the approach, we conducted extensive tests using various top-tier cross-domain recommendation algorithms. The results were promising: our model outperformed the competition, registering a preliminary performance surge of 5% over conventional benchmarks.

As an NSF National Research Trainee Fellow at the Center of Adaptive Systems of Brain and Body Interaction, I led an interdisciplinary team to develop a web application aimed at helping incarcerated individuals access reentry services. I managed the development using React for the front-end, Node/Express for the back-end, and MySQL for the database. Additionally, I initiated key partnerships with the Director of DC Jail and the Vice President of American Prison Data Systems (APDS) to trial the application on APDS tablets, facilitating real-world implementation.

Taught 60 person labs twice a week for two semesters, while creating programming assignments for the students to work on. Held office hours every week to help students understand programming concepts, and worked on grading labs and programming assignments for over 100 students.

Working as a teaching assistant for Data Mining Course at GMU. Assisted students on machine learning algorithms, and help them understand machine learning and data science concepts (Classification and Clustering Algorithms). Help students work with KNN, Neural Networks, K-Means clustering, Decision Trees, and held office hours every week.

Worked on application for hospitals to keep track of patient and staff data in hospitals. I developed API’s to retrieve patient information, doctor information, and track call logs. I used Node JS for developing API’s, and React JS for frontend development. The data for the application was stored in AWS S3, and it was deployed on an EC2 Instance. In addition, a CI/CD pipeline was set in place for developing the web application.