whoami
I am Ozel Yilmazel, a senior year Computer Science undergraduate at UMass Amherst. I have a deep passion for machine learning and natural language processing. Currently I am working at Center for Data Science, my past project was the iNatAtor, my current project is developing a geospatial annotation tool for Buzzards Bay, this project is NSF funded. I have interned at Center for Data Science at UMass, where I contribute to the research paper SINR by improving GeoModels on iNaturalist through the development of a web-based annotation tool called iNatAtor. I took the lead on fine tuning GeoModels with annotation data. In addition, I developed a data extractor which transformed saved annotations on iNatAtor into data usable for fine tuning GeoModels. I also took part in developing the front-end of the application, integrated a Postgres database to manage annotations, and containerized the application to assist in deployment. This project has advanced the development of open-source projects, iNaturalist and SINR. Additionally, I have a strong interest in web development and am continuously learning new frameworks to expand my skill set.
Goals
I am actively applying to graduate schools for a PhD program. My main research interest is to develop language models with potentially different inference architectures. Current models are exceptionally well, but have become very big, and lately all the innovation we have seen has been just increasing model sizes to get better results. During my PhD I hope to bring answers to questions about what alternatives are there, how can we teach more human-like representation of language to models (and would this make the models smaller?). The potential future that may come up from these answers open up incredible research opportunities.
I am also seeking a summer internship before I start my (potential) graduate year at UMass.
You can reach me at ozel@yilmazel.com.
Education
University of Massachusetts Amherst | Bachelor's of Science in Computer Science
- GPA: 4.0
- Coursework: Data Structures CS187, Programming Methodologies CS220, Computer Systems Principles CS230, Reasoning Under Uncertainty CS240, Introduction to Computation CS250, Algorithms CS311, Web Applications CS326, Practice and Applications of Data Management CS345, Principles of Data Science CS348, Artificial Intelligence CS383, Machine Learning CS389, Applications of NLP CS485
- Current Coursework: Information Systems CS445, Search Engines CS446, Theory and Practice of Software Engineering CS520, Game Programming CS576
Experience
Research fellow | DataCore | September 2024 - Present
Center for Data Science at UMass Amherst
- Currently working on Buzzards Bay GeoSpatial Interview Annotation Tool. This is another annotation tool similar to iNatAtor that will collect annotations from folks in the Buzzards Bay region. This project will potentially incorporate information retrieval processes to search up particular annotations that relate to a query. This project is NSF funded.
- I have worked on improving iNatAtor, our annotation platform for iNaturalist.
- Enhanced iNatAtor by implementing open backlog tickets from the summer program and patched some bugs related to frontend.
- Revamped the frontend application by transporting existing code to use Mantine material library.
Research fellow | iNaturalist | Summer 2024
Data Science for the Common Good | Center for Data Science at UMass Amherst
- Contributed to the research paper SINR by developing a web based annotation tool called iNatAtor to improve the GeoModels used by thousands on iNaturalist.
- Developed an extension of SINR, a fine tuning program that used annotation data gathered from the iNatAtor application.
- Theorized and developed new loss functions to handle new objectives and accommodate new labels.
- Contributed future research on GeoModels by writing a report on fine tuning and problems faced with the new annotation data.
- Integrated a Postgres database to manage the saved annotations in the application.
- Engineered frontend rendering optimizations that improved the overall application performance.
- The project contributed to developments of open source projects: iNaturalist and SINR.
- We have a poster.
Software Engineer & NLP Engineer | Sociail (pronounced "social") | June 2023 - April 2024
- Developed backend services that managed API requests to our LLM instances.
- Engineered context-size management solutions for LLMs which reduced our costs.
- Built a semantic analysis model that classified use inputs in order to call the appropriate generative model to satisfy their request.
Undergraduate Course Assistant | September 2022 - May 2024
UMass Amherst
Projects
CatGAN
CatGAN is a modified CycleGAN model developed in pytorch. CatGAN achieves domain translation of images, Humans to Cat. The translations are learned through features extracted from convolutions, therefore each human has a unique cat transformation, and every cat has a unique human translation. CycleGAN paper originally recognizes the limitations of geometric transformations, this project shows it is possible.- Implemented my own CycleGAN model to train unpaired image transformations from Humans to Cats and back to Humans.
- Contributed to the research by showing geometric transformations are achievable by modifying model architecture and prioritizing a loss objective.
- GitHub repository, which also includes a demo video.
- Utilized Google Colab to deliver the project.
Twitter Bot Classification
This project is a research report that focuses on classifying human and bot accounts on Twitter, using only textual data (username and tweet). In this report we explore a variety of approaches to semantic analysis: shallow models, transformer ensembles, and few-shot learning with LLMs. Each approach gets more complex than the previous one and achieves higher scores. Our results using transformer ensembles put us among the best results discussed in the TwiBot-20 paper. Our report provides semantic analysis that is competitive to other bot classification techniques that use the full metadata of accounts. We believe that when our approaches are combined with ambitious metadata classifiers it could achieve higher classification scores.- Engineered semantic analysis solutions using shallow models: Naive Bayes, Logistic Regression, Decision Trees, Random Forests.
- Fine tuned BERT and Twi-BERT to classify tweet data and trained a random tree model to classify based on username and tweet data, all three models were combined to create an ensemble logistic regression model.
- Slide for the project that shows results and architecture. It is an animated slide, you use presentation mode to view animations.
- Utilized Google Colab to deliver the project.
TuneLink
TuneLink is a music-only sharing platform, similar to TikTok but with strictly music sharing purpose. I have been working with 4 other friends to bring this project to life, we expect to be finalized by December 2024.- Developed the backend services using express.js and node.js.
- Tested developed API using Jest.
- Integrated MongoDB for document and file storage.
OverMath
A math game based on OverCooked video-game. Made in Unity. I have been working with 4 other friends to develop a funny and challenging puzzle game that stimulates mathematical thinking in a different way. We expect to finish by December 2024.- Implemented character controllers.
- Implemented adversarial AI in the game.
StackBuildIO
StackBuildIO is an AI powered tech-stack builder/generator for those who are not familiar with what to use on their projects. You can enter a prompt to describe your project and it will refer you to a tech-stack. You can save and share your tech-stack with other users on the platform, and can make edits any time you want.- Developed a full stack web application using MEHN stack (MongoDB, Express.js, HTML, Node.js).
- Fine tuned GPT-3.5 to assist users better and return correct JSON format of responses.
- GitHub repository.
- Youtube Demo
Skills
A list of skills I have obtained from both my coursework at UMass and my work experiences.
- Python: My go to language in any data science related task. I have used python in CatGAN, Twitter Bot Detection, and Fine-Tuning GeoModels.
- Javascript: My go to language when I have to build a web application. I have used JS in iNatAtor, Buzzards Bay, StackBuildIO, TuneLink.
- SQL: I have to use SQL regularly because of my coursework but also in my projects at Center for Data Science
- Java, C: Other complementary languages I have experience in.
- HTML, CSS: What you see right now.
- PyTorch, TensorFlow: I am more experienced in PyTorch and it is my preferred ML/NN library. I used them both in CatGAN and Twitter Bot Detection. I used PyTorch significantly in Fine-Tuning GeoModels
- NLTK: I have used this toolkit in every NLP project including Twitter Bot Detection.
- LaTeX: I regularly had to deliver homeworks and research reports.
- React, Node.js, Express.js, Mantine: I have used these in every JS project.
- GitHub: I have used GitHub to both version control my own projects and on work projects.
- Postman: I have used Postman to write test suites for our APIs.
- Docker: I have used docker to containerize iNatAtor, my projects, and Buzzards Bay for deployment and sharing.
- Google Colab: I desperately needed GPU. Used in every personal ML project. For Fine-Tuning GeoModels, I have used the Unity Computer Cluster at UMass.
- Google Cloud: I use it deploy simple container applications.
Extras
Impact of Software Development in Startups
- I gave a presentation software development in startups, especially at a stage when there is no clear tech stack.
- Slides
- I wrote a whitepaper on ethical considerations of face recognition technology
- Paper
- I try to publish my new findings and thinkings on medium. Right now I have talked about Ethics in CS, a novel way to store context windows in LLMs, and the effect of pre-processing data.
- My Medium Wall