Saehee Eom

GT EI-Lab · Prof. Mark Riedl · Ongoing

Research on training data attribution for large language models. Enriched pre-training corpus with 576 topic × format bins via stratified and proportional sampling, investigating data quality signals and provenance to improve LLM pre-training transparency.

~5B tokens curated · 576 topic × format bins

GT Sonification Lab · Ongoing

Agentic AI framework with autonomous agents orchestrating multimodal data collection — audio, facial, and behavioral signals — combined with LLM-based clinical reasoning for continuous mental health monitoring.

Emotion AI · multimodal time series · autonomous agent coordination

SNU SKI-ML Lab · Under revision

Co-authored research on energy model-based controlled text generation for black-box LLMs, handling multiple constraints simultaneously — toxicity, logical consistency, formality — while preserving fluency. Fine-tuned via SFT using synthetic datasets.

+23% BERTScore · −22% perplexity · +31% constraint satisfaction across LLaMA, Phi, Qwen

SNU · B.A. Thesis · December 2024

Quantified the effects of syntactic complexity on LLM performance across QA, text completion, and EN–KO translation using 8 linguistically motivated metrics. Identified POS divergence as the key bottleneck and demonstrated prompt engineering strategies to mitigate degradation.

In-context learning, kNN example selection, instruction tuning

SNU MIPA Lab · 2024

Adversarial defense system against diffusion-based deepfakes — injects imperceptible noise into training photos to degrade unauthorized AI-generated identity recreation and protect against identity misuse.

−2% generated image quality · −3% facial similarity degradation

SNU · 2nd place / 40 teams

Two-stage detection pipeline combining object detection and OCR for intelligent transportation systems. Won 2nd place out of 40 teams in competition.

87.60% OCR accuracy · 0.72 mean IoU · 81% end-to-end detection

Georgia Tech · 2025

Hybrid search engine combining structured SQL queries with FAISS vector similarity search, enabling ANN-based retrieval and relational filtering on large-scale text corpora. Adaptive query strategies dynamically tune search scope based on filter selectivity.

Hyundai Home Shopping · 2022

Ensemble model combining LightGBM gradient boosting and SASRec sequential recommendation to optimize customer targeting for marketing campaigns. Deployed on AWS for real-time predictions with sentiment analysis across 10K+ reviews.

+42% purchase rate · +14% campaign engagement lift

Growth Hackers · SNU · 2022

Hybrid CatBoost + LSTM forecasting for call center volume with an interactive Streamlit dashboard. Enabled managers to optimize workforce allocation and reduce operational costs.

MAE < 100 · 29% lower monthly error vs baseline

LBOX · EMNLP 2022 Workshop

South Korea's first legal data analytics platform — end-to-end development visualizing sentencing patterns and enhancing accessibility through information extraction pipelines.

4× engagement from non-legal users · 161% YoY sales increase

Georgia Tech · 2025

Web-based traffic prediction dashboard for Atlanta's I-285 with 330,000+ data points. Features historical/forecast mode switching and animated timeline visualization with geospatial rendering via Leaflet.js.

75% compression ratio · <5MB · 1–3s initial load

Georgia Tech · 2025

SQL database engine built from scratch with a Lark-based parser, supporting DDL/DML operations. Integrated collaborative filtering for book recommendations.

<1s on 100K+ rows · +18% Precision@10 vs baseline