Sihun Lim

ECE @ Seoul National University. Systems, GPUs, and agentic AI for hardware.

Email · LinkedIn

About

B.S. in Electrical & Computer Engineering at Seoul National University (expected Feb 2027).
Interned on the System Software Team at Rebellions, building memory profiling for their AI chip.
Studied hypervisor and agentic system security at the SNU Computer Security Lab with Prof. Byoungyoung Lee.
Lately: agents that design chips, and memory for agents.

Projects

• Agentic FPGA backend optimization (2026) — an autoresearch meta-harness that optimizes Vivado designs from DCP checkpoints and learns from execution history and sparse human feedback. 1st place, FPL'26 Agentic FPGA Backend Optimization Competition (AMD).
• Agentic chip backend automation (2026) — MCP tools over Synopsys ICC2 and Siemens Calibre so an agent can run Tcl, read DRC/LVS violations, and iterate on physical-design and ECO fixes. Validated on TinyRISCV and a TPU systolic array.
• Temporal Cortex (2026) — a graph-based long-term memory system for LLM agents with hybrid graph + text retrieval and a Python SDK. Showcased at AGI House Long Horizon Build Day.
• GPU memory profiling @ Rebellions (2025–26) — tools to visualize OS heap usage and HBM bandwidth at 1 ms granularity during inference.
• Dynamic GPU allocation (2025) — a VS Code extension and web interface for allocating GPUs only when you actually need them. Second-round interview, South Park Commons.
• StyleGAN2 acceleration (2025) — 20× faster than the naive GPU implementation using CUDA. Accelerator Programming Summer School, Advanced Track.
• Graph-based data preprocessing (2024) — a VS Code extension for building data preprocessing pipelines by wiring Python functions together as nodes. 1st place, Air Force Start-up Challenge.

Skills

C, C++, CUDA, Python, JavaScript.
GPU programming and inference serving (vLLM, NVIDIA Dynamo), Kubernetes, AWS.