Information & Communication Engineering student focused on Edge AI, on-device inference, and Physical AI systems.
I am interested in a practical systems question: how can AI models run efficiently and reliably under real device constraints? I explore that question through embedded devices, controlled measurement, and kernel-level profiling.
- Edge / On-device AI — model quantization, GPU offload, memory use, latency, and throughput
- Physical AI — the interaction between perception, compute resources, and real-world action
- AI systems profiling — separating end-to-end behavior from individual kernel behavior
- IoT systems — sensing and control with Raspberry Pi and embedded peripherals
I am currently working on an LG Aimers hackathon project. The repository will be published once the implementation and results are ready to share.
This team project will explore multi-sensor verification and quantization robustness on Raspberry Pi or Jetson. The repository will remain private during development. Now we are searching references and studying.
I profiled Qwen3.5-2B inference with llama.cpp, NVIDIA Nsight Systems, and Nsight Compute.
| Comparison | Observed result |
|---|---|
| Whole-prefill throughput, B=384 → B=512 | 916.24 ± 32.67 → 992.18 ± 38.02 token/s (+8.3%) |
| Selected BF16 Tensor Core GEMM kernel | 9.636 ± 0.538 → 8.564 ± 0.230 TFLOP/s (−11.1%) |
Takeaway: a slower sampled kernel does not necessarily mean lower whole-model throughput. I report kernel-level and end-to-end measurements separately and control workload, warm-up, clocks, and repetitions.
📌 Edge AI profiling repository
Built and evaluated an offline Korean meeting-transcription pipeline combining speech recognition, local LLM correction, and speech synthesis on a resource-constrained edge device.
📌 Meeting transcription repository
| Project | What it shows | Stack |
|---|---|---|
| Edge AI on Jetson | Completed Jetson Orin Nano inference experiments with end-to-end and kernel-level profiling | Jetson, llama.cpp, Nsight Systems / Compute |
| Offline Meeting Transcription | Completed edge pipeline for Korean speech recognition, local LLM correction, and speech synthesis | Jetson, Whisper, local LLM, TTS |
| IoT Systems coursework | Raspberry Pi sensing and control with PIR, DHT11, ultrasonic, gas, light, soil-moisture, motors, and simple vision exercises | C, Python, Raspberry Pi, OpenCV |
| Job Buddy | Team repository for an AI-assisted resume and cover-letter web service | FastAPI, OpenAI / Gemini APIs, Spring Boot, React, MySQL |
- Languages: Python, C/C++, Java, Kotlin
- Edge & AI systems: Jetson Orin Nano, Raspberry Pi,
llama.cpp, CUDA, model quantization - Profiling: NVIDIA Nsight Systems, Nsight Compute, controlled benchmarking
- Application development: FastAPI, Spring Boot, React, MySQL, Android / Jetpack Compose
- Supporting hands-on learning as an IoT Systems teaching assistant
- Currently studying Physical AI / Cyber-Physical AI, efficient LLM inference, and experimental methodology for AI systems
- Turning course experiments into reproducible repositories with clear setup, measurements, and limitations
- I am also curious about teaching someone.
Question → controlled setup → end-to-end measurement → kernel inspection → interpretation → limitations
I prefer evidence-backed conclusions and explicitly separate what was measured from what is inferred.