Technical Stack技術工具
The foundational tools and frameworks I leverage to build resilient, high-performance AI architectures. 用來建構穩健、高效能 AI 架構的基石工具與框架。
Backend & APIs後端與 API
FastAPI, Uvicorn, SQLAlchemy
LLM & Agent OrchestrationLLM 與代理協同
DeepSeek Tool Use, OpenAI Codex CLI, Ollama / Qwen
Vision & Document AI視覺與文件 AI
Qwen2-VL OCR, PyTorch, PyMuPDF
Data & Retrieval資料與檢索
SQLite (WAL), Brave Search API, JSON/JSONL Pipelines
Enterprise Integration企業系統整合
Odoo ERP (JSON-RPC), KGI Securities API, pCloud
Automation & Messaging自動化與通知
APScheduler, Windows Task Scheduler, Discord Bots
Document & Media Generation文件與媒體生成
Playwright → PDF, Jinja2, Kokoro TTS
DevOps & DeploymentDevOps 與部署
Docker Compose, Nginx + Let's Encrypt, uv
Judgment Calls判斷/取捨
Three calls that come up again and again across these projects — and the reasoning behind them. 這些專案裡反覆出現的三個判斷,以及背後的理由。
Not the Newest — the Best Fit不是最新的,是最適合的
Simple questions don't go through the expensive deep-research pipeline; when keyword search is accurate enough, I don't stand up a vector database; a monitor running 24/7 uses a local model instead of burning API fees. I pick technology by cost, latency, and scale — not by what's trending. 簡單問題不進昂貴的深度研究流程;關鍵字搜尋夠準,就不架向量資料庫;24 小時跑的監控用本地模型,不燒 API 錢。選技術看成本、延遲、規模,不看流行。
How People Use It Comes First先想人怎麼用,再想怎麼做
When you're stuck on a book, the hardest part is turning "I don't get this" into a clear question — so the AI Library lets you drop a note right on the passage or diagram. The AI knows which page that is, and answers from that page's actual content. 看書卡住的時候,最難的是把「我哪裡不懂」問清楚——所以 AI 書庫讓你直接在段落或圖表上貼一張筆記,AI 知道那是書裡哪一頁,就用那頁的內容回答。
Technical Skill, Plus Alignment技術之外,也要對齊人
At Surglasses I took part in product-direction discussions and cross-department coordination; as a freelancer I work with clients directly, no PM in between. Explaining trade-offs in plain language, taking feedback, and adjusting course is as much part of the job as writing code. 在骨王生技參與過產品方向討論與跨部門協調;接案時直接面對客戶,中間沒有 PM。把取捨用平實的話講清楚、接回饋、調方向,跟寫程式一樣是工作的一部分。