M.S. in Computer Science @ National Taiwan University · B.S. in CS @ NYCU
Systems software · LLM serving & inference systems · Computer vision · Backend infrastructure
- Industry experience at Trend Micro, the Institute for Information Industry (III), and Academia Sinica: evaluation frameworks for LLM-agent tool interfaces, LLM training pipelines on DGX clusters (DeepSpeed/LoRA), multi-agent LLM systems, and edge AI deployment in C++.
- Computer vision from classical operators to deployed systems: object detection and multi-object tracking, domain adaptation for cross-domain species classification, and a real-time tracking pipeline served as low-latency HLS.
- Comfortable across the stack, from CUDA kernels, MPI/AVX-512 optimization, and SSD firmware simulation to concurrent network servers and containerized Linux infrastructure.
Most of my recent work (internships, thesis research) lives in private repositories; the projects below are a curated public subset.
- EvolVE: Evolutionary Search for LLM-based Verilog Generation and Optimization Hsin, W.*, Deng, R.*, Hsieh, Y.*, Huang, E. & Hung, S. · arXiv:2601.18067, 2026. (*equal contribution) [paper]
- Structured Testbench Generation for LLM-Driven HDL Design and Verification-Oriented Data Curation Huang, E., Kao, Y., Deng, R., et al. · arXiv:2606.12983, 2026. [paper]
- Exploring relationships among soundscape perception, spatiotemporal sound characteristics, and personal traits through social media Chan, T., Wu, B., Lee, Y., Lee, P. & Deng, R. · Noise Mapping, 10(1), 2023. [paper]
| Project | Description | Tech |
|---|---|---|
| hpc-kernel-optimization | Step-by-step kernel optimization: attention kernel from 241s to 3.2s (75.6×) on a 4-node cluster | C, CUDA, OpenMP, MPI, AVX-512 |
| p-language-compiler | Full compiler for the P language: scanner → parser → AST → semantic analysis → RISC-V codegen | C/C++, flex, bison, RISC-V |
| ssd-ftl-labs | Instrumenting SSD internals in the MQSim simulator: FTL mapping-table hit rate, GC, and wear leveling (incl. an upstream bug fix) | C++, Python |
| unix-systems-programming | A ptrace + Capstone instruction-level debugger, and an LD_PRELOAD syscall-interception sandbox | C/C++, ptrace |
| mips-cpu-verilog | MIPS processor built bottom-up in Verilog: structural ALU, single-cycle datapath, and 5-stage pipeline | Verilog |
| verilog-handwritten-digit-recognition | MNIST neural-network inference implemented entirely in Verilog (FSM-driven), with an interactive Python GUI | Verilog, Python |
| Project | Description | Tech |
|---|---|---|
| network-programming | Concurrent network servers from scratch: piped remote shell, multi-user server (select / fork + shared memory), HTTP+CGI, SOCKS4 proxy | C++, BSD sockets, boost.asio |
| onos-virtual-router | SDN virtual BGP router as an ONOS application: intent-based forwarding, Quagga peering, ARP gateway, DHCP relay | Java, ONOS, BGP, Mininet |
| spark-bigdata-analytics | Big-data analytics on multi-million-row datasets with PySpark, Spark MLlib, and a self-provisioned GCP Dataproc cluster | PySpark, Hadoop, GCP |
| freebsd-sysadmin-toolkit | FreeBSD sysadmin tools: a ZFS snapshot-backup CLI, an rc.d FTP upload-policy monitor, and a JSON container decoder | Shell, FreeBSD |
| smart-fridge | Multi-user smart fridge on Raspberry Pi: face recognition, Gemini-based item detection, per-user inventory via Telegram bot | Python, Raspberry Pi |
| Project | Description | Tech |
|---|---|---|
| video-tracking-and-streaming | Classification → YOLOX detection with an SE-attention neck → ByteTrack MOT → a real-time tracker served to the browser as low-latency HLS at 30 FPS | PyTorch, OpenCV, ffmpeg, Flask |
| uav-autopilot-vision | Autonomous Tello flight from vision alone: camera calibration, ArUco pose + PID visual servoing, a fine-tuned YOLOv7-tiny detector, and grid-based line following | Python, OpenCV, PyTorch |
| image-processing-and-colorization | Histogram equalization/specification, Gaussian convolution, bicubic interpolation and Butterworth notch filtering implemented from the arithmetic up, applied to archival photo restoration | NumPy, OpenCV |
| Project | Description | Tech |
|---|---|---|
| multimodal-lora-gapro | Combining LoRA-GA initialization with LoRA-Pro gradient correction on multimodal LLMs (LLaVA, Qwen-VL); unified training/eval framework with DeepSpeed | Python, PyTorch, DeepSpeed |
| nlp-and-llm-finetuning | Four NLP tasks from BiLSTM-attention to LoRA fine-tuning of Vicuna-7B; using one LLM to filter another's context lifts accuracy 0.714 → 0.802 | Python, PyTorch, FastChat, LoRA |
| ai-capstone | Search, inference & ML agents: hand-written MCTS and a CNF-resolution Minesweeper solver | Python |
| decision-tree-implementation | CART classification tree from scratch with linear-time incremental Gini split search | Python, NumPy |
| Suika-Game-DRL-Final | Deep RL agents for the Suika game across four observation/action difficulty levels | Python, PyTorch |
| AI_Project | Icon style transfer with CycleGAN and pix2pix implemented from the papers | Python, PyTorch |
- EvolVE: evolutionary search for LLM-based Verilog generation (arXiv:2601.18067); co-first author.
- AnimeFrameBot: Go API server with fuzz/mutation testing; contributed the
/frameand/uploadAPI endpoints, fuzz tests and coverage, and the Telegram bot functions. - ICRTL-Benchmark: industrial-level RTL design challenges for evaluating LLM-based RTL generation; research collaboration (follow-up to EvolVE).
| Award | Detail |
|---|---|
| Global Winner, NASA International Space Apps Challenge (2022) | Global Connection Award, Top 10 of 5,327 teams · 3D_ISS_Tracker |
| Team of Excellence (Top 5), Taiwan Presidential Hackathon (2024) | "Quiet Tracker", edge-AI urban noise governance |
| 1st Place, Data De-identification and Re-identification Competition, Academia Sinica (2023) | |
| 3rd Place, Meichu Hackathon, NXP Enterprise Award (2023) | |
| 1st Place, Meichu Hackathon, STMicroelectronics Enterprise Award (2022) |
| Category | Technologies |
|---|---|
| Languages | C/C++, Python |
| Systems & Infrastructure | Linux (Arch/Ubuntu), Docker, Kubernetes, Ansible, PostgreSQL, Git/GitLab, GitHub Actions |
| Parallel & GPU Computing | CUDA, OpenMP, MPI, SIMD (AVX-512) |
| AI & GenAI | PyTorch, vLLM, SGLang, DeepSpeed, LangGraph, LLM fine-tuning (LoRA), RAG, multi-agent systems, MCP tools, LLM evaluation (LLM-as-judge) |