My Projects

Participated as a data scientist mentor at the Duke Critical/Acute Care Datathon 2025, a collaborative two-day event connecting critical care clinicians with data scientists. The datathon focused on developing pragmatic data-driven models using de-identified critical care electronic health record datasets (MIMIC, eICU) with the theme 'Data Science in Critical/Acute Care.' Teams were organized with half data scientists and half clinicians to develop new projects in 36 hours, from problem identification to abstract completion. As a mentor, I guided interdisciplinary teams in applying machine learning and data science techniques to improve the care of critically ill and injured patients, fostering collaboration between healthcare professionals and data scientists.

I am honoured to be one of the 10 recipients worldwide of the prestigious 2024 Global Women in Technology Scholarship by Palantir Technologies. This scholarship recognises women who are pursuing careers in technology and demonstrates exceptional academic achievement, leadership, and a passion for computer science and engineering. As a recipient, I have been awarded a grant of $7,000 to support my education and will participate in professional development workshops aimed at preparing me for a successful career in technology. This recognition motivates me to continue striving for excellence and to contribute significantly to the tech industry.

Our paper, "Towards Optimizing and Evaluating a Retrieval Augmented QA Chatbot using LLMs with Human-in-the-Loop," has been published in the ACL Proceedings of the Fifth Workshop on Data Science with Human-in-the-Loop (DaSH 2024). Our work was also awarded Best Paper at NAACL 2024. This research focuses on optimizing a Retrieval-Augmented Generation (RAG) pipeline, enhancing performance through a human-in-the-loop approach. Collaborating with SAP SE, we developed an HR support chatbot to address employee inquiries. This involved integrating domain experts, improving response quality, exploring alternative retrieval methods, and employing state-of-the-art evaluation metrics.

During AaltoAI's Hackathon at Aalto University in Finland, my partner and I secured first place in the Microsoft GenAI Challenge, competing against eight other teams. Our project, SPARROW (Smart Proposal Assistant for RFP Response and Offer Workflow), aimed to automate the Request For Proposals (RFP) process using real-world product data. We leveraged Langchain and OpenAI Embeddings in Pinecone, with GPT-4 generating RFPs in various formats. A RAG pipeline facilitated product retrieval and matching, while GPT-4 finalized the selections and Microsoft's Phi-3 composed the emails. This automated solution enhances time efficiency by streamlining the response process for RFPs, utilizing crucial product data from the company's database.

Large Language Models have found application nowadays in many fields such as repetitive tasks including HR support. We worked with the domain experts of SAP to develop an RAG HR support chatbot as an efficient and effective tool for addressing employee inquiries. By enhancing the LLM driven chatbot’s response quality and exploring alternative retrieval methods, we have created a more efficient, scalable, and flexible tool for HR professionals to address employee inquiries effectively. To evaluate and further enhance the prompts for the used LLMs as well as to improve the retriever, we explored many evaluation metrics and compared them with human evaluation. We constrasted state-of-the-art, reference-free evaluation metrics against traditional reference-based metrics to discern a deeper understanding of text quality.

In an NLP course, we advanced IVAN, a German news voice assistant, using the RASA framework for natural language processing. After evaluating 17 models, we integrated a fine-tuned T5 model, achieving results comparable to ChatGPT. This integration not only significantly improved IVAN's ability to understand and generate human-like responses but also enhanced its efficiency in delivering tailored news content. We enhanced personalization with a sophisticated data model that tailors content to user preferences, and used automated ChatGPT testing for quality assurance. Additionally, this rigorous testing process ensured that IVAN maintained a high standard of reliability and user satisfaction. We also improved news search functionality with Levenshtein distance and word2vec embeddings for accurate entity recognition, thereby enabling more precise and relevant news delivery to our users. You can find a short demo of the final project by pressing this card.

In my research for an NLP course, I studied the enhancement of conversational agents, focusing on task-based and social types to improve human-machine interaction. I evaluated various models, from early chatbots to advanced systems like ChatGPT, and examined their applications in fields like customer support and healthcare. A significant part of my work involved analyzing architectural methods for building these agents, including pipeline and end-to-end frameworks, and response generation techniques. This research offered insights into the evolving landscape of conversational agents and identified future research directions. Additionally, the investigation emphasized the critical role of natural language understanding and generation in creating more effective and engaging conversational experiences.

In my Bachelor thesis, I delve into the intricate role of NLP in the design and development of algorithms for chatbots, utilizing a comprehensive range of older deep learning models such as the sophisticated Sequence-to-Sequence frameworks and the Transformer architectures, which have significantly shaped the current landscape of AI. By assessing seven models' dialogue performance with BLEU metrics and qualitative reviews, the research highlights the inadequacy of translation metrics for dialogue systems, advocating for metrics that better reflect human judgment. This study suggests directions for future chatbot development towards more natural interactions, marking a significant evolution in chatbot technology. The paper’s code is accessible for further exploration and development.

In developing my portfolio website, I focused on sharpening my Next.js and UI/UX design skills, aiming to craft a user-centric online showcase. This platform highlights my key projects and personal interests, offering visitors an intuitive and engaging journey through my professional landscape.

Our team of five in a software development course created "Robo Rally" a Java/JavaFX strategic board game for 2-6 players over 12 years. It leverages object-oriented and event-driven programming for a lively game environment with double-sided boards, robot figures, and cards. Players aim to navigate robots through a factory, hitting checkpoints before others. We used Java for the game's complex logic and JavaFX for the UI, enabling appealing visuals, animated movements, and a simple control panel for programming robot actions. The project, exploring game design and user experience, resulted in an engaging game. We deployed an AI Agent in case the player plays solo as well. The link below shows a short demo of the game.