I design AI systems that transform complex real-world problems into structured, evidence-grounded workflows. My current interests include LLM agents, retrieval-augmented generation, knowledge graphs, financial research automation, policy intelligence, and applied machine learning.
I am an M.S. candidate in Digital Analytics at Yonsei University, researching multi-agent systems, various types of RAG, and knowledge graphs.
Semantic Tabular-to-Image Conversion and Contrastive Learning for Lightweight Intrusion Detection
A lightweight intrusion-detection framework that transforms tabular network-traffic data into RGB images for CNN-based representation learning in IoT edge environments.
My contributions
- Designed the framework using SHAP-based feature selection and LLM-categorized Vortex Feature Positioning
- Combined a depthwise separable CNN encoder with contrastive pretraining
- Evaluated the model across six IDS and IoT benchmark datasets
- Achieved average binary Accuracy/F1 of 96.4% and multiclass Accuracy/F1 of 90.2%
- Reduced the model to 9,662 parameters and 38.3 KB
Tech: Python PyTorch SHAP LVFP CNN Contrastive Learning IoT Security
Publication: Journal of the Korea Society of Computer and Information, Vol. 31, No. 3, March 2026
Collaborative research project. The original repository is maintained by a research collaborator.
Point-in-Time, Evidence-Grounded Multi-Agent Financial Research System
A multi-agent system that combines financial statements, disclosures, news, and market data to generate traceable investment judgments and equity research reports.
My contributions
- Designed the workflow connecting financial, news, market, strategy, and writer agents
- Defined primary-evidence, secondary-context, and structured-output contracts
- Designed point-in-time controls to prevent information published after the analysis date from entering the report
- Separated source evidence from LLM-generated interpretation and controlled duplicate evidence aggregation
- Developed an explainable report pipeline with provenance tracking and validation logic
Tech: Python LLM Agents Agent Orchestration Evidence-Grounded AI Financial Analysis Structured Outputs
Collaborative research project. The original repository is maintained by a research collaborator.
Multi-Agent Policy Issue Discovery System — KISTEP Phase 1
A policy-intelligence pipeline that discovers Korean science and technology innovation issues from large-scale news data and converts them into structured issue cards.
Key components
- Collected Korea-related news through GDELT BigQuery
- Applied source filtering, multi-stage article crawling, preprocessing, and three-tier chunking
- Built a custom RAPTOR retrieval structure using recursive summarization and hierarchical search
- Designed STEEP specialist agents for social, technological, economic, environmental, and policy perspectives
- Added Gate 0 quality validation and regeneration for final issue cards
Tech: Python LangGraph RAG RAPTOR GDELT BigQuery ChromaDB Multi-Agent Systems
Repository is currently private because the project contains research-specific implementation and data configurations.
Two-Stage K-Means Clustering and P-Median Location Optimization
A data-driven policy analysis project that identifies care-vulnerable neighborhoods in Seoul and proposes optimized locations for community care service hubs.
Key contributions
- Integrated elderly population, disability, transportation, facility, terrain, and accessibility data across 432 administrative districts
- Constructed a unified analysis dataset with 56 indicators
- Applied two-stage K-Means clustering to identify 148 vulnerable neighborhoods and classify them into seven vulnerability types
- Applied population-weighted P-median optimization to evaluate single and multiple service-hub scenarios
- Proposed Banghwa 3-dong and Suseo-dong as priority candidates for additional community care hubs
- Built a reproducible Python pipeline with automated tests and GitHub Actions
Tech: Python pandas GeoPandas scikit-learn
K-Means P-Median Geospatial Analytics pytest GitHub Actions
Ontology-Based Knowledge Graph-Augmented SLMs for Leakage-Controlled Bug Triage
A graduate research project investigating how domain-specific ontology and knowledge-graph evidence can improve small language models for bug component prediction and assignee recommendation.
Research scope
- Analyze bug-report fields, label distributions, and time-based train/validation/test splits
- Define a bug-triage ontology covering BugReport, Product, Component, Assignee, Term, and historical relationships
- Construct a historical knowledge base with text retrieval and an ontology-based knowledge graph
- Compare Vector RAG, general GraphRAG, and a proposed domain-specific BugTriage-RAG method
- Evaluate component prediction using Accuracy and F1
- Evaluate assignee recommendation using Recall@K and MRR
- Apply history-only retrieval and leakage-control rules throughout the experiment
Tech: Python Small Language Models Ontology Knowledge Graph GraphRAG Vector RAG Neo4j Bug Triage
Research and implementation are in progress. The repository will be published after the experimental structure is stabilized.
Aug 2023 – Nov 2023 · Seongnam, South Korea
- Reproduced navigation, communication, and functional failures of autonomous mobile robots in a ROS2-based environment
- Classified defects and translated findings into actionable R&D improvement tasks
- Debugged new RCS features and validated ROS2–RCS integration
- Contributed to a 30% reduction in defect rates, a 20% shorter testing cycle, and more than 100 documented bug reports
May 2023 – Aug 2023 · Pangyo, South Korea
- Analyzed global fandom data and click-through-rate trends to support product-marketing strategy
- Conducted country-level conversion-rate and market-segmentation analyses
- Supported the global expansion of Learn! KOREAN with BTS and related initiatives associated with approximately 750,000 cumulative global sales
M.S. Candidate in Digital Analytics
Mar 2025 – Feb 2027 (Expected) · GPA: 3.98 / 4.3
B.S. in Statistics
Minors in Mathematics and Economics · Dec 2024 · GPA: 3.53 / 4.0
- DataFest — 1st Place, Best in Show
- Dean’s List — 6 of 8 semesters
Python · SQL · R · SAS · PostgreSQL · Neo4j · BigQuery · Git · GitHub · GeoPandas · Shapely
LLM Agents · Multi-Agent Systems · LangGraph · RAG · RAPTOR · GraphRAG · Agent Orchestration · Evidence-Grounded Generation · Structured Output Design
PyTorch · SHAP · Feature Selection · CNN · Depthwise Separable CNN · Contrastive Learning · K-Means · P-Median · K-Means Clustering · P-Median Optimization · Location-Allocation
Statistical Analysis · Financial Analysis · Market Segmentation · CTR Analysis · Conversion Analysis · Geospatial Analytics · Policy and News Data Analysis · Model Evaluation
ROS2 · RCS Integration Testing · System Debugging · Defect Reproduction · Failure Classification · Test Scenario Design · Bug Reporting
- Email: hoini0920@gmail.com
- LinkedIn: linkedin.com/in/hoinlee
- GitHub: github.com/TaylorLee99