Building Trustworthy AI via
Formal Verification & Mathematical Rigor
"I study the discrepancy between an LLM's internal motives and external behaviors (Alignment Faking) and build trustworthy AI systems by mathematically formalizing these phenomena."
Jae-Hyun Baek
M.S. in Mathematics, Sogang Univ.
M.S. in Mathematics
2024.03 - 2026.02 (Expected)Sogang University
GPA: 4.1 / 4.3 (Top 3%)
Advisor: Prof. Jon-Lark Kim
Focus: Coding Theory, AI Safety, Optimization
B.S. in Mathematics
2022.03 - 2024.02Sogang University
Transfer Student | GPA: 3.8 / 4.3
1. Formal Methods as a Weapon
Unlike typical engineers, I approach AI Safety from a mathematical foundation. I don't just find vulnerabilities; I aim to provide Formal Guarantees and explain why alignment failures occur structurally.
2. Global Communication
Proven ability to present complex research in English at international venues (ISIS 2025 Award). Ready to collaborate with global leaders like Anthropic & DeepMind.
3. Engineering to Research
Capable of building the full evaluation pipeline (EntropyMath). I bridge the gap between abstract theoretical safety and practical, scalable engineering.
Goal: Research methodology to induce Alignment Faking via Red Teaming.
Taxonomy & Setup: Define "Strategic Compliance" triggers in finance/medical domains.
Red Teaming Experiments: Develop automated jailbreaking prompts that target "Sycohphancy" and "Faking".
Evaluation & Paper Draft: Quantify the "Faking Rate" and draft a submission for NeurIPS/ICML 2026.
Team Leader & Eng. Lead
DeepFountain Corp. (2025.03 - Present)
Led the development of AI Agent Platforms (SaaS). Orchestrated Ruby on Rails + Python hybrid architecture for scalable RAG systems.
Research Intern
DeepHelix (2025.01 - 2025.02)
Developed Multi-modal LLM pipelines for medical education assistants.
AI & Research
Web Engineering
Selected Projects
SolEvolve (Master's Thesis)
Autonomous algorithm discovery system using LLM and Evolutionary computation. Rediscovered optimal codes ($[22,11,6]$) via Black-box optimization.
EntropyMath (AI Safety Eval)
Leaderboard for evaluating Agentic Tool Use, designed to solve Data Contamination issues in LLM benchmarks.
SOGAMBOT.com
University-wide AI Chatbot and RAG solution for Sogang University.
MekaNet (Medical AI)
WSI-based Tiny Object Detection using SSL architecture. Focused on medical image analysis.