Technical Case Studies技術案例研究

AI Solutions
Engineering
AI 解決方案
工程

The AI projects I've actually built. Each one started from a real problem and went all the way to running in production — click in to see the thinking and the trade-offs behind it. 這裡放實際做過的 AI 專案。每個都從一個真實問題開始,一路做到能上線穩定運作;點進去可以看到當初的想法和取捨。

Autonomous research agent dashboard
Agentic Research自主研究代理
description Case Study案例研究

Autonomous Deep Research Engine自主深度研究引擎

Good research isn't about reading everything — it's about finding the sources that actually matter and turning them into something readable, without burning tokens on searches that don't move the needle. 好的研究不是把所有東西都讀過一遍——是精準找到真正有效的資料來源,整理成好讀的摘要,不把 token 浪費在無效的搜尋上。

Tech Stack技術工具

Python FastAPI OpenAI Codex CLI SQLite + JWT

Solution Impact解決方案效益

  • Simple questions get an answer in seconds instead of waiting through a research process they don't need簡單問題幾秒內就有答案,不用跑一次它根本不需要的完整研究流程
  • Real-time progress means never staring at a blank screen during a multi-minute research run即時進度讓使用者不會盯著空白畫面等好幾分鐘
  • Multiple people can research at once without one person's query slowing down anyone else's多人可以同時做研究,不會因為一個人在查資料就拖慢其他人
Equity research synthesis pipeline
Grounded Synthesis可溯源合成
open_in_new Live Demo線上 Demo

Automated Evening Report自動化晚報

A trader can't read 100–200 broker reports by hand and still react to the market in time — a full day isn't enough. This pipeline distills them overnight so researchers and traders can catch what's moved by the time the market opens. 交易員沒辦法手動看完 100–200 份券商報告還來得及對市場變化做反應——一整天都不夠用。這套流程在夜間就先蒸餾完,讓研究員跟交易員一開盤就能掌握真正變動的重點。

Tech Stack技術工具

Python DeepSeek Tool Use Brave Search

Solution Impact解決方案效益

  • Turns what would be hours of manual reading into a report that's ready before the market opens把原本要花好幾個小時人工讀完的工作,變成開盤前就準備好的一份報告
  • Every hard fact is graded by source reliability, so readers know at a glance what to trust more每項硬事實都標記了來源可信度,讀者一眼就知道哪些可以更信任
  • Tracks how a story evolves across days, catching shifts a single day's snapshot would miss追蹤一則新聞跨日如何演變,抓到單日快照會漏掉的變化
AI library knowledge distillation interface
Multi-Agent Pipeline多代理流程
description Case Study案例研究

AI Library & Knowledge DistillationAI 書庫與知識融煉

A whole book doesn't fit in a single ChatGPT or Claude conversation — you can't just paste it in and ask it to explain. This system OCRs, chunks, and distills the material first, so the AI can actually help you learn from something too large to hand it directly. 一整本書塞不進一次 ChatGPT 或 Claude 的對話裡——不能直接貼上去要它幫你整理。這套系統先做 OCR、拆解、蒸餾,讓 AI 真正能處理這種直接丟給它會爆掉的龐大內容,幫助你有效利用 AI 學習。

Tech Stack技術工具

React FastAPI Qwen2-VL OCR Kokoro TTS

Solution Impact解決方案效益

  • Makes a 200-page technical book something you can actually query and get grounded answers from, not just skim讓一本兩百頁的硬核技術書變成真的可以提問、拿到有根據答案的東西,而不是只能翻閱
  • Distilled Skill modules mean what you learned from one book can be reused in other AI workflows, not locked in your head蒸餾出來的 Skill 模組讓你從一本書學到的東西,可以重複用在其他 AI 工作流程裡,不是只留在腦中
  • Answers cite the actual page they came from, so you can go verify it yourself檢索答案根據實際頁面內容生成並附出處,可以自己回去核對
Real-time market alert agent
Event-Driven Agent事件驅動代理
description Case Study案例研究

Futures Market Alert Agent期貨市場警報代理

By the time a market-moving headline gets picked up, price and positioning may have already reacted — so the only edge left is closing the gap between when news breaks and when a trader sees it. This agent watches 15+ sources around the clock so a human doesn't have to. 一則會影響市場的新聞被注意到時,價格與籌碼往往已經先反應了——剩下能拚的就是縮短「新聞發生」到「交易員看到」中間的落差。這個代理 24 小時盯著 15+ 個新聞源,人不用自己盯盤。

Tech Stack技術工具

Python asyncio Ollama / Qwen FastAPI SQLAlchemy + SQLite

Solution Impact解決方案效益

  • Removes the need for a human to manually monitor 15+ news sources around the clock不用再有人 24 小時盯著 15+ 個新聞源
  • A single misleading headline can't trigger a false alarm — the weighted score requires multiple signals to agree單一則誤導性標題觸發不了假警報——加權分數需要多個訊號一起同意才會發
  • The same story breaking across 5 outlets triggers one alert, not five, so the signal doesn't get lost in noise同一則新聞被 5 家媒體報導,也只觸發一次警報,訊號不會被雜訊淹沒
Enterprise internal portal
Enterprise Portal企業內部平台
open_in_new Live Demo線上 Demo

Enterprise Internal Portal & ERP企業內部入口平台與 ERP

Traders and staff at the firm were logging into a different tool for every task — trade records in one place, clock-in/leave in another, training materials somewhere else. This portal puts all of it — trading, training, knowledge base, AI research, and the company's Odoo ERP — behind a single login. 公司裡的交易員與員工,每件事都要登入不同系統——交易紀錄一個地方、打卡請假另一個地方、教學資料又是別的地方。這套入口平台把交易、教學、知識庫、AI 研究,跟公司的 Odoo ERP,全部整合到同一個登入後面。

Tech Stack技術工具

FastAPI Vanilla JS (JSX) Odoo ERP (JSON-RPC) KGI Securities API APScheduler Docker + Nginx

Solution Impact解決方案效益

  • One login replaces separately logging into trading, HR, training, and research systems一次登入取代原本要分別登入交易、人事、教學、研究好幾套系統
  • Clock-in records feed straight into payroll and generate the PDF payslip — HR isn't reconciling it by hand打卡紀錄直接串到算薪資、產生 PDF 薪資單,不用人資手動對帳
  • Futures positions sync in real time into the same platform — no need to open a separate broker terminal to double-check期貨部位資料即時同步進同一個平台,不用另外開券商系統核對