PurpleHound BAS AI Agent
In development- Designed and implemented a tool-based AI agent for natural-language use of BAS capabilities, including scenario generation, simulation execution, result analysis, and alternative suggestions.
- Validated action ordering and implemented the product execution API to connect scenario generation with execution.
Security knowledge & attack-path exploration
- Developed a graph- and ontology-based security knowledge system.
- Implemented security-knowledge retrieval, relationship exploration, and an attack-path engine.
PurpleHound-Embed-v1
- Fine-tuned a 0.6B embedding model with synthetic queries to address retrieval errors in mixed Korean–English security queries.
- With fine-tuning and BM25 hybrid retrieval combined, improved Recall@5 from 0.710 to 0.887 on a mixed Korean–English query evaluation.
Internal document search
- Proposed and built an internal search project to find scattered development documents using natural language.
- Used document structure and metadata for retrieval, and deployed a Slack Bot integration within the existing workflow.
Haehoon Seo
I’m interested in turning AI’s diverse possibilities into practical applications and expanding where it can be applied. At the same time, I explore ways to verify AI outputs so they can be trusted. At 78ResearchLab, I work as an ML engineer at the intersection of AI and security, developing AI agents, retrieval systems, and knowledge systems. I studied statistics at the University of Seoul. Outside of work, I enjoy CrossFit.
Contact
Careers
Projects
PRed
I planned and developed this app to solve the friction I experienced logging my own CrossFit training. It turns photos or text into CrossFit and Hyrox workout logs, using accumulated training history for AI coaching and PR tracking.
Research
Papers & presentationsALIBI: Adversarial Legitimacy Injection in Binary Input against LLM Malware Analyzers
arXiv:2609.19722 · Preprint ·
ARES: Auxiliary Range Expansion for Outlier Synthesis
arXiv:2501.06442 · Preprint ·
EDAL: Energy-driven Active Learning
The Korean Statistical Society Fall Conference · Conference poster · 2024
Race-Robust Deep Learning-Based Facial Expression Classification Technique
한국방송미디어공학회 학술발표대회 논문집 · Conference proceedings · 2024
Previously
- MLA Lab, University of Seoul (2024.01–2025.01)
Research in out-of-distribution detection and active learning.
- Korea Electronics Technology Institute (2024.07–2024.08)
Research on fairness in facial expression classification.
- Optim Lab, University of Seoul (2023.08–2023.12)
Worked on a domain-specific chatbot pipeline.
Education
University of Seoul
Department of Statistics
2022–
Northumbria University
Computer Science · Exchange student
Jan–Jun 2025
Awards
Top Excellence Award, Data Analysis Competition
University of Seoul
Encouragement Award, Poster Paper Awards
The Korean Statistical Society
Excellence Award, Student Competency Poster Presentation
University of Seoul
Academic Excellence Award · 3 times
Seoul Mayor’s Scholarship
Ranked first in the department
Certifications
ADsP