B.S. Candidate @ KAIST, School of Computing and Nuclear & Quantum Engineering, Minor in Mathematical Sciences
Working to become a researcher.
A professor once told me that writing a paper means diverging first, then converging everything into one clean, linear narrative. This is my attempt to visualize that idea with a reverse-time SDE I built: pure random noise (left) is gradually denoised into structure, while the sampling trajectory (right) wanders the whole probability landscape before converging to a single point.
The common thread is HW-SW co-design: shaping hardware and software together, not either in isolation.
- ML Systems Architecture: architecture design for ML workloads such as LLMs and 3D Gaussian Splatting, with particular interest in state management and offloading.
- Compiler Design for ML: the software half of co-design, compilers that map ML workloads onto the hardware they run on.
- Verifiable Computer Architecture: hardware whose correctness is proven rather than benchmarked, since edge cases slip past even the best test suites.
- FPGA Accelerators: domain-specific accelerators for near-storage processing and in-storage processing.
- GPU Systems: topology, communication, and parallelism across multi-node clusters, on both the scale-up and scale-out axes.
- Emerging Technologies: bringing new substrates such as CXL, HBM, HBF (high-bandwidth flash), and quantum computing into conventional computer architecture.
Beyond research
- LLM Hallucination: verification-driven ideas, in the spirit of Lean-style formal checking, to keep model claims verifiable.
- Knowledge Graphs: converge diverged ideas into structure, then re-diverge the structure to expand them (NodeGraph).
- Dynamic-Capacity Neural Networks: networks whose parameter count grows and shrinks during training, the way 3DGS densifies and prunes.
- Reactor Physics and Transmutation Lab @ KAIST (Winter 2024, Advisor: Prof. Yonghee Kim): Ran large-scale Monte Carlo neutron transport simulations with OpenMC and analyzed their computational, memory, and parallel-scalability characteristics.
- Nuclear I&C and Autonomous Operation Lab @ KAIST (Winter 2025, Advisor: Prof. Jonghyun Kim): Resolved a GPU-underutilization bottleneck in a heterogeneous CPU-GPU reinforcement-learning pipeline by decoupling policy optimization from a CPU-bound, closed-source simulator, reverse-engineering its process memory, and bridging a Windows VM and a GPU server with a TCP/Docker-based distributed infrastructure.
- Computer Architecture and Memory Systems Laboratory @ KAIST (Summer 2026, Advisor: Prof. Myoungsoo Jung): Completed an intensive seminar on computer architecture and memory systems, centered on the internals of SSDs.
- Computer Architecture and Memory Systems Laboratory @ KAIST (Fall 2026, Advisor: Prof. Myoungsoo Jung): Individual research, ongoing.
- VRAIL (Co-author, equal contribution): Vectorized Reward-based Attribution for Interpretable Learning. arXiv:2506.16014, 2025.
- SHIFT (First author): Sigmoid-Based Heuristic Invertible Fitness-Landscape Transformation for Accelerating SBST. arXiv:2604.09171, 2025.
- [CS420] KECC: A full C-to-RISC-V compiler written in Rust with an SSA-based IR (Mem2reg, GVN, DCE), matching Clang -O1 in cycle-level gem5 evaluation (1.009x).
- [NQE625] Nuclear System Code (Steady, Transient): A multi-channel 1-D FVM nuclear system code in MATLAB, validated against the MARS code across 10 physical quantities.
- [Hobby] NodeGraph: A VS Code extension that turns papers into verifiable knowledge graphs, built around working with AI agents.




