AI Pipeline Toolkit - Grace Weng

The Real Problem

真正的問題

The obvious problem was volume: 6 platforms (LinkedIn, CakeResume, Yourator, 104, Meet.jobs, company career pages), different formats, different update cadences. But the real problem was signal. Most of my time was spent on roles that weren't a good fit. I had no systematic way to know: which roles actually match what I'm good at? Where should I invest my limited time?

在 6+ 平台間追蹤機會意味著在瀏覽器分頁間切換、搞不清楚哪些已經看過、錯過時效性項目。每個平台的格式不同、更新頻率不同，也沒辦法交叉比對訊號。我需要一個能聚合、分析、並凸顯重點的系統。

The System

系統

I built a triage and prioritization system. It aggregates roles from 6 sources daily, uses AI to score each one against my actual experience, and surfaces the ones worth my time. The dashboard gives me a morning brief (what's new, what needs follow-up), a pipeline view (where is each opportunity), and a career coach (interview prep from my own stories). The result: I went from spending most of my time browsing to spending most of my time preparing for the right conversations.

我打造了一個自動化 pipeline，每日從 6+ 平台抓取資料，用 AI 評分與排序，透過本地 web dashboard 呈現所有內容。系統包含優化建議、跨訊號模式分析，以及自動化的 email 摘要。

How It Works

運作方式

Morning Brief showing daily triage with new jobs, follow-ups, and conversion funnel

Morning Brief: daily triage with new matches, follow-ups due, and funnel stats

每日簡報：新配對、待追蹤、以及漏斗統計

Review Queue with AI match scoring at 95%

Review Queue: AI scores each role against your profile with match reasoning

審核佇列：AI 根據你的背景為每個職缺評分並說明原因

Pipeline view tracking applications across companies

Pipeline: track every opportunity from discovery to interview with next actions

Pipeline：追蹤每個機會從發現到面試的進度與下一步

Career Coach for interview prep with STAR story building

Career Coach: interview prep that builds STAR stories from your actual experience

Career Coach：根據你的實際經歷建構 STAR 故事的面試準備

Architecture

架構設計

6+ Platform

Scrapers

6+ 平台爬蟲

→

Data

Normalization

資料

規範化

→

AI Scoring

(Claude API)

AI 評分

(Claude API)

→

Dashboard +

Pipeline

Dashboard +

Pipeline

Key Decisions & Learnings

關鍵決策與學習

Why Playwright for Scraping?為什麼用 Playwright 做爬蟲？

— Most platforms require JavaScript rendering and authentication. Playwright handles both, plus provides screenshot capabilities for debugging. The headless browser approach also handles anti-bot protections better than raw HTTP requests. — 大多數平台需要 JavaScript 渲染和驗證。Playwright 兩者兼顧，還提供截圖功能方便除錯。無頭瀏覽器的方式也比原始 HTTP 請求更能應對反爬蟲機制。

AI Scoring FrameworkAI 評分框架

— Built a multi-criteria scoring system using Claude API that evaluates items across dimensions like relevance, quality, and timing. The framework is configurable — criteria and weights can be adjusted without code changes via a YAML config. — 用 Claude API 建立多維度評分系統，從相關性、品質、時效等面向評估項目。框架可配置——透過 YAML 設定檔即可調整評分標準和權重，無需改程式碼。

Local-First ArchitectureLocal-First 架構

— Deliberately kept everything local — no cloud database, no SaaS dependencies. Data stays on my machine, the dashboard runs on localhost, and the cron job handles scheduling. This keeps costs at zero and data private. — 刻意維持全本地架構——沒有雲端資料庫，不依賴 SaaS。資料留在本機，dashboard 在 localhost 運行，cron job 處理排程。成本歸零，資料也保有隱私。

Note: This pipeline was originally built for Taiwan's job market, where platforms like CakeResume, Yourator, and 104 are dominant. Job search strategies differ significantly across markets. In the US and Europe, referrals and warm introductions often matter more than applications. The tool solves the aggregation and triage problem well, but the broader strategy would look different outside Taiwan.

備註：此 pipeline 最初為台灣求職市場打造，主要整合 CakeResume、Yourator、104 等本地平台。不同市場的求職策略差異很大，歐美市場更重視內推與中間人介紹。這個工具很好地解決了資訊聚合與篩選的問題，但在台灣以外的市場，整體求職策略會不同。

Job Search Command Center

求職指揮中心