• Elevated a voice deepfake detection pipeline by upgrading nueral network architectures (AASIST) and integrating 20K+ diverse datasets (synthesized, cloned, recorded), enhancing model generalization and cross-domain scalability to date.• Designed and deployed a scalable dataset splitting logic with seamless integration of SAMO-specific loss functions, preserving data integrity across the entire modeling pipeline.• Engineered a robust augmentation pipeline using RawBoost to address extreme class imbalance and incorporated error analysis workflows to refine model accuracy, achieving a current EER reduction to 0.05 during validation.At AirLab, Department of Electrical and Computer Engineering, University of Rochester, we are advancing the SAMO (Speaker Attractor Multi-Center One-Class Learning) framework to enhance its effectiveness in voice anti-spoofing and deepfake detection. Our enhancements include integrating advanced machine learning technologies to improve model adaptability and robustness.Our efforts extend to broadening the application of core concepts to address additional speech attributes, which will enhance detection capabilities. We are also rigorously testing these innovations on a variety of datasets to ensure robustness and applicability across diverse environments.

Emergency Response & Resource Optimization for City of Rochester Fire Department • Led the student team to analyze 1.6M+ incident records using time-series models (Prophet, SARIMAX) and geospatial tools (Folium, GeoPandas), forecasting incidents and optimizing resources for 500 personnel managing 50K annual dispatches.• Delivered predictive models with (85% R² accuracy), evaluated fire stations placemebts, and recommended reallocation strategies for 19 vehicles, through comprehensive reports and html interactive maps.• Secured recognition with media features on the University of Rochester News Center.

• Synthesized a 14K audio dataset by integrating and fine-tuning 18 generative AI and neural network models (Text-to-Speech, voice cloning), enabling the development of a robust deepfake detection model with enhanced accuracy and robustness.• Developed and executed a data collection strategy with 28 key parameters, optimizing parameter distributions to collect 400K metadata values, reducing dataset bias, and enhancing generalization.As a Data Scientist Intern at IngenID, a leading provider of voice biometric solutions handling over 24M+ transactions annually, I contributed to advancing voice biometric authentication with a focus on deepfake detection. My primary responsibility was the creation of a comprehensive speech dataset for anti-spoofing applications, leveraging cutting-edge deep learning models such as Tacotron, FastSpeech, Glow-TTS, VITS, and HiFi-GAN.In collaboration with the Audio Information Research (AIR) Lab at the University of Rochester, I implemented advanced Text-to-Speech (TTS) and voice cloning models to gather and analyze thousands of voice samples from diverse open sources and commercial APIs. I also designed and implemented robust metadata collection strategies, conducting detailed investigations into acoustic and metadata features to optimize IngenID’s classification algorithms.This project played a critical role in enhancing the accuracy of IngenID's deepfake detection systems, strengthening the security of its voice biometric solutions across multiple industries in response to the growing threats posed by synthetic audio technology.

As a Data Scientist:• Conducted and automated NLP analysis of 50K+ messages from Korea's #1 user-ranked chat community in the category.• Increased purchases by 25% and eliminated manual analysis of 300K+ messages annually, saving $6,500 per year.• Built ETL pipelines and engineered features for a recommendation system, decreasing dropout rates by 10%.As a co-founder:• Spearheaded end-to-end startup operations, from ideation to investor engagement, directly contributing to a strategy towards an educational platform for art students, resulting in a user base expansion across South Korea and the US.• Successfully raised $3M in seed funding by leveraging data-driven storytelling and impactful visualizations in pitch decks.

E-commerce retailer of Korean food for U.S. market (#1 in category on Amazon US) • Analyzed 7,800+ product reviews using NLP to uncover customer behavior patterns, boosting repeat purchases by 13% and driving the launch of two new products targeting loyal customers.• Optimized Amazon ad campaigns for a $4M annual sales product line by refining keywords and budgets, increasing RoAS by 17%, resulting in an additional $680K in annual revenue.• Improved project outcomes through ongoing 1:1 mentorship by Nvidia Korea's ex-CEO.