About Me
I'm Sheng-Wei (Roger) Fan, a CS student at NTHU. I work across the stack — front end, backend services, and the infrastructure underneath. I've also done ML/CV research, and I work with LLM agents daily — directing and reviewing what they produce, in hackathon projects and at work. I built this site and I run it, down to the mail server and DNS. The Neovim setup below is the one I work in every day.
Dev Environment
My everyday terminal environment — Neovim, Tmux, and Zsh, customized in Lua and shell and Catppuccin-themed throughout. It's tuned to handle everything from coding to LaTeX and notes, and one script sets it all up on a fresh macOS or Linux machine.
- •One-command bootstrap:
initialization.shinstalls Neovim, Nerd + CJK fonts, Tmux, and CLI tools (ripgrep, fzf, bat), then symlinks every config into place - •Neovim tuned into a full IDE: 60+ lazy-loaded plugins covering LSP (Mason), completion, linting, formatting, Telescope, Treesitter, and Copilot — backed by 950+ lines of keymaps
- •More than an editor: LaTeX with VimTeX, LeetCode with leetcode.nvim, and Obsidian notes with live Markdown rendering — all without leaving Neovim
- •Git in-editor: inline hunks with gitsigns and a full lazygit TUI
- •Tmux + Zsh: custom-styled status bar, session restore, and seamless pane jumps between Neovim splits and Tmux (vim-tmux-navigator); Zsh with Powerlevel10k, autosuggestions, and syntax highlighting
- •Layered config: a tracked universal layer plus a gitignored per-machine layer, keeping machine-specific tool paths (Flutter, gcloud, uv…) out of the shared dotfiles
Education
National Tsing Hua University
Bachelor of Science in Computer Science
- •GPA: 4.24/4.3 (Top 3% in Department of Computer Science) cert
Key Courses
Experience
R&D Intern
Synology (NAS Group · C2 Division · CAT Team)
Jul 2026 - Aug 2026New Taipei, Taiwan- •Built the Vue front end for a license upgrade feature serving customers, resellers, and distributors — delivered through weekly iteration with PM, UX/UI, and QA, and a strict code review process.
- •Extracted SMS delivery from a monolithic backend into a standalone Go service, with the data model and delivery semantics settled with the team before implementation — asynchronous send, transactional enqueue, idempotent retries, multi-pod concurrency, signed callbacks, and state derived from records rather than held in memory.
Receiving the certificate of internship from the VP
With my manager and mentor
Summer Research Intern
Institute of Information Science, Academia Sinica
Jul 2025 - Aug 2025Taipei, Taiwan- •Built head-wise cross-attention heatmaps that surfaced failure patterns, pinpointing and fixing flaws in a video-editing diffusion model.
- •Designed experiments with custom object-crossing datasets, exposing the limitations of VideoGrain.
Teaching Assistant
Introduction to Programming (II) & Linear Algebra, NTHU
Feb 2025 - Jun 2026Hsinchu, Taiwan- •Built a C++/Allegro game template that let students apply OOP concepts in a full project.
- •Designed and graded assignments, projects, and exams for 70+ students across three English-taught semesters.
Awards
Top Prize, 2026 AIWave: Taiwan GenAI Hackathon | NT$100,000
1111 Job Bank Challenge · powered by AWS
Built SkillWeave, an AWS job search ranking system. Amazon Bedrock extracts and normalizes skills from 1.2M postings into a weighted Neptune skill graph; OpenSearch (BM25 + embedding kNN) cuts the corpus to 200 candidates; a LambdaMART re-ranker trained on real search logs returns the top 20, each with an evidence trace. A geo graph of Taiwan's administrative districts expands zero-result searches to adjacent and parent areas instead of dropping the location filter. Query understanding ran on batched Bedrock calls at 10:1 compression with a two-tier cache (offline precompute + runtime LRU), keeping LLM parsing off the request critical path.
College of EECS Outstanding Student Scholarship | NT$120,000
National Tsing Hua University
Merit award given to the college's top-ranked students, selected through a review of both academic and extracurricular performance.
3rd Place, 2026 TSMC IT CareerHack — AAID Challenge
TSMC IT CareerHack
Addressed the challenge of single-image outdoor light estimation by conducting comprehensive error analysis and 25+ ablation experiments. Identified and mitigated the model's reliance on global geometric anchors (horizon lines) by introducing Spatial Attention (CBAM) and optimizing backbone architectures. The final EfficientNet-ResNet hybrid model achieved 0.9867 AUC on the SynthStreet dataset.
1st Place, NSF HDR Scientific-MOOD Hackathon in Taiwan
US National Science Foundation (NSF) Harnessing the Data Revolution (HDR)
Developed ML models to predict drought severity (SPEI index) from NEON ground-beetle specimen images. The challenge emphasizes out-of-distribution generalization, requiring models to perform reliably across unseen ecological sites with different climates and species distributions.
Academic Achievements Award
National Tsing Hua University