With 25+ years in embedded systems, I've always liked getting to the bottom of things: how does this hardware actually work? And when something goes wrong, where exactly did it go wrong?

Much of that time has been spent bringing up boards that nobody has touched before — from U-Boot and drivers to Secure Boot, and the tricky points where hardware, software and FPGA don't quite agree. Many of those problems only reveal the answer when you trace them all the way down.

Lately, I've been bringing that same engineering mindset to AI and open-source projects. I don't rush to call myself an expert in something new. I'd rather build it, test it in practice, and when it doesn't behave as expected, find out why.

• Embedded Linux: Kernel development, U-Boot, Driver optimization, RootFS (Yocto/Buildroot)
• Board Bring-up & HW/SW Integration: Expert on NXP (LX2160), Marvell (ARM), and Intel platforms
• Security: Secure Boot implementation (Cisco Trust Anchor), 5G-ORAN (Genevisio)
• High-Speed Interfaces: 25G SerDes, PTP/SyncE optimization
• Networking Products: Full lifecycle delivery — concept → prototyping → mass production → field support
• Cross-Team Collaboration: Bridging gaps between SW, HW, and FPGA teams; mentoring through code reviews and architectural guidance

Notable background: Delivered mission-critical networking equipment at Cisco (ASR1K, ISR, Network Modules).

In my spare time I like turning ideas into tools that actually get used, and picking up new technology by building with it. From embedded diagnostics and Edge AI to AI applications and automation, each project is a run at making the idea real, testing it, and making it better.

Bringing the diagnostic mindset from real product development to the M5Stack CoreS3.

Built on C, ESP-IDF and Clean Architecture: a modular embedded diagnostic tool combining sensor monitoring, an interactive UART console, and power and thermal monitoring. It takes the maintainability, testability and layered design I care about in product work and applies them to a personal hardware project.

An offline storyteller (M5Stack CoreS3)

Running a small language model on an ESP32-S3, where resources are tight. A 28.9M-parameter model is deployed to the M5Stack CoreS3, with Per-Layer Embeddings putting the 25M-parameter lookup table in Flash, measured at roughly 9.5 tokens per second. The project also covers bilingual story generation, Chinese rendering on the LCD, and temperature sampling — an exploration of what Edge AI on a microcontroller can look like.

Ever pasted something into a chatbot and then thought — "wait, what happens to this once it's sent?"

That question is what this project is about. Before data reaches the LLM, it checks for sensitive information, replaces personal data, and drops columns that aren't needed; when the AI answers, the response is checked too. Through testing and evaluation, it explores how to keep a bit more control over your data while still using AI.

Ever wondered whether you could build your own AI working environment — one that keeps sensitive data safe even when you use cloud GPUs, without paying a monthly AI subscription?

That's why I built Self-Cloud-Agent-Lab. From local models and RAG to MCP tool integration, and now exploring cloud GPUs, it's a step-by-step build of an AI environment I control. The goal isn't just to get AI running, but to explore the balance between data safety, compute cost and real performance — so AI doesn't have to depend on an off-the-shelf paid service.

• 🌟 AI-Beginner-GitHub-Compass — React, Node.js, Gemini and a Python crawler collect trending GitHub projects and recommend where to start based on your interests and learning direction, with search, bookmarking and local data management.
• 🔧 JobSpy — Python scraping of embedded roles from LinkedIn, Indeed, Seek and Jora in Melbourne (forked from cullenwatson/JobSpy), plus skill-match scoring, career-track classification and a Kanban board to keep the job hunt organised.
• 🏢 AI-Company — Exploring how multiple AI agents can divide the work, breaking the software development process into 18 distinct roles and tasks.
• 🧠 AI-MindMeld — Four AI models take different roles in the same discussion, and a chairperson pulls the views together into a conclusion; pure frontend, and playable offline with no API key.
• ⚙️ Claude-Config — Claude Code conventions and workflow organised by project type (C/Embedded, Kotlin/Android, Python AI, TypeScript/Web), covering architecture design, task breakdown, code review and verification.

• Advanced AI integration patterns
• Full-stack web development
• Cloud deployment strategies
• Modern embedded tooling and automation

• Open-source AI projects
• GitHub ecosystem tools
• Web scraping and data analysis
• Developer productivity tools
• Embedded systems tooling

• 🤖 Artificial Intelligence & Machine Learning
• 🛠️ Embedded Systems & Hardware-Software Integration
• 🔐 Secure Boot & Platform Security
• 🌐 Web Development & Full-Stack Applications
• 📊 Data Analysis & Web Scraping
• 💼 Business Automation

• LinkedIn: ian-chang-56136479
• GitHub: @ian0318git
• Location: Melbourne, Australia
• Open to: Senior Embedded Systems roles, collaboration on interesting projects

• 🎯 I believe in the power of AI to democratize software development
• 🚀 I don't just write code — I identify systemic bottlenecks and bridge communication gaps between cross-functional teams
• 🌍 Always excited to explore new technologies and frameworks
• 💪 Love contributing to open-source and learning from the community
• 📖 Passionate about mentoring and raising team standards through robust code reviews
• 🔄 Seen products through the entire lifecycle — from concept to mass production and field support

擁有 25 年以上嵌入式系統開發經驗，我一直喜歡追根究柢：硬體究竟怎麼運作？出了問題，又是在哪個環節出了差錯？

這些年，我做得最多的事之一，就是讓一塊沒人碰過的板子真正運作起來。從 U-Boot、驅動程式到 Secure Boot，再到硬體、軟體與 FPGA 彼此兜不起來的地方——很多問題，得一路追到最底層才有答案。

最近，我也把這套工程思維帶進 AI 與開源專案。對新技術，我不急著宣稱自己精通；我更在意的是把它做出來、實際驗證，並在它不如預期時，找出真正的原因。

• Embedded Linux: Kernel 開發、U-Boot、驅動程式優化、RootFS (Yocto/Buildroot)
• Board Bring-up 與 HW/SW 整合: 精通 NXP (LX2160)、Marvell (ARM) 和 Intel 平台
• 安全機制: Secure Boot 實作 (Cisco Trust Anchor)、5G-ORAN (Genevisio)
• 高速介面: 25G SerDes、PTP/SyncE 優化
• 網路產品: 完整生命週期交付——概念 → 原型 → 量產 → 現場支援
• 跨團隊協作: 橋接 SW、HW 與 FPGA 團隊之間的溝通差距，透過 Code Review 與架構指導提升團隊水準

重要經歷: 在 Cisco 參與關鍵網路設備開發 (ASR1K、ISR、Network Modules)。

業餘時間，我喜歡把想法變成實際可用的工具，也藉由動手實作探索新技術。從嵌入式診斷、Edge AI 到 AI 應用與自動化，每個專案都是一次把想法做出來、實際測試並持續改進的過程。

把真實產品開發中的診斷思維，帶進 M5Stack CoreS3。

以 C、ESP-IDF 與 Clean Architecture 為基礎，開發模組化的嵌入式診斷工具，整合感測器監測、UART 互動式 Console，以及電源與溫度監控。把過去產品開發中重視的可維護性、可測試性與分層設計，應用在自己的硬體專案上。

離線說故事的說書人（M5Stack CoreS3）

讓資源有限的 ESP32-S3 也能執行小型語言模型。將 28.9M 參數的模型部署到 M5Stack CoreS3，運用 Per-Layer Embeddings 把其中 25M 參數的大表放進 Flash，實測推論速度約每秒 9.5 個 token。專案也整合中英文故事生成、中文 LCD 顯示與溫度取樣，探索微控制器執行 Edge AI 的可能性。

你有沒有過這種經驗：把資料貼進聊天機器人後，才突然想到——「等一下，這些資料送出去之後，會發生什麼事？」

這個專案就是想解決這個問題。在資料送進 LLM 之前，先檢查敏感資訊、替換個人資料，並移除不必要的欄位；收到 AI 回應後，再進行檢查。透過測試與評估，探索如何在使用 AI 的同時，對資料多一點掌控。

有沒有想過，能不能打造一套自己的 AI 工作環境？即使使用雲端 GPU，也能兼顧敏感資料的安全，不必每個月花錢訂閱 AI 服務？

這就是我打造 Self-Cloud-Agent-Lab 的原因。從本機模型、RAG 到 MCP 工具整合，再探索雲端 GPU，逐步搭建一套自己能掌控的 AI 環境。我想做的不只是把 AI 跑起來，而是探索如何兼顧資料安全、運算成本與實際效能，讓 AI 不一定得依賴現成的付費服務。

• 🌟 AI-Beginner-GitHub-Compass——結合 React、Node.js、Gemini 與 Python 爬蟲，蒐集熱門 GitHub 專案，依照使用者的興趣與學習方向提供探索建議，並支援搜尋、收藏與本地資料管理。
• 🔧 JobSpy——以 Python 蒐集 LinkedIn、Indeed、Seek、Jora 上墨爾本的嵌入式職缺（fork 自 cullenwatson/JobSpy），加上技術匹配評分、職涯方向分類與 Kanban 管理，讓求職流程更有系統。
• 🏢 AI-Company——探索如何讓多個 AI Agent 分工合作，把軟體開發流程拆解成 18 個不同角色與任務。
• 🧠 AI-MindMeld——讓四個不同角色的 AI 參與同一場討論，再由主持者整理觀點、歸納結論；純前端，免 API Key 也能離線試玩。
• ⚙️ Claude-Config——依專案類型（C/Embedded、Kotlin/Android、Python AI、TypeScript/Web）整理的 Claude Code 規範與工作流程，涵蓋架構設計、任務拆解、程式碼審查與驗證。

• 高級 AI 集成模式
• 全棧 Web 開發
• 雲端部署策略
• 現代嵌入式工具與自動化

• 開源 AI 項目
• GitHub 生態工具
• 網頁爬蟲和數據分析
• 開發者生產力工具
• 嵌入式系統工具鏈

• 🤖 人工智慧與機器學習
• 🛠️ 嵌入式系統與軟硬體整合
• 🔐 Secure Boot 與平台安全
• 🌐 Web 開發與全棧應用
• 📊 數據分析與網頁爬蟲
• 💼 業務自動化

• LinkedIn: ian-chang-56136479
• GitHub: @ian0318git
• 地點: 澳洲墨爾本
• 開放機會: Senior Embedded Systems 職位，也歡迎有趣的項目合作

• 🎯 我相信 AI 的力量能夠民主化軟體開發
• 🚀 我不只是寫程式——我善於識別系統性瓶頸，並在跨職能團隊之間搭建溝通橋樑
• 🌍 對探索新技術和框架充滿熱情
• 💪 喜歡為開源做貢獻並向社區學習
• 📖 熱衷於透過 Code Review 和架構指導提升團隊標準
• 🔄 經歷過產品從概念到量產再到現場支援的完整生命週期