# Kos 的代表案例

# Kos's case studies

Canonical: https://www.aikos.cc/cases/

## 支付與錢包系統

技術負責人・帶領 13 人以上團隊・2024–2026

高併發交易下，回呼重送或網路重試都可能讓同一筆款項被處理兩次；會員餘額與資金流水必須隨時對得上。

- 支付冪等、分散式鎖與餘額一致性機制，同一筆交易只入帳一次
- 補償與重試機制處理失敗交易，搭配風控稽核
- 以 Kafka、Redis、MongoDB 建立事件驅動架構
- 自動對帳流程，加上交易監控與告警
- 以重送、併發與補償情境的測試驗證上述機制

- 不重複扣款: 冪等與分散式鎖
- 自動對帳: 補償與稽核流程

公司名稱與營運數據暫不公開。可公開的是測試情境：回呼重送、併發請求與補償，沒有另附線上交易量。

## 新聞採集與全文搜尋平台

傲創科技・資深工程師・2019–2024

每接一個新聞網站就要寫一套爬蟲，約需 2 天；全文搜尋回應約需 2 秒。

- 以 Hyperf、Swoole 建立高併發採集服務，RabbitMQ 調度與削峰
- 可配置式採集框架，規則存在 MongoDB，新站點改用設定接入
- Elasticsearch 全文搜尋，Canal 將 MySQL 即時同步到索引
- Redis 與 APCu 多層快取，降低資料庫負載

- 約 2 天 → 半天內: 新站點接入時間
- 約 2 秒 → 300 毫秒內: 搜尋回應時間
- 十萬級: 每日處理新聞資料
- 數千篇: 尖峰每分鐘採集量

搜尋時間比較的是全文搜尋上線前與上線後的回應，不是公開的百分位測試，也沒有留下查詢樣本數。

## 履歷問答助手

本站・規則工具＋只看履歷的模型

訪客會自由提問。模型若編造公司、數字或日期，履歷就不可信；模型逾時或失敗時，頁面也不能沒有答案。

- 頁面先把已知問題交給履歷工具：公司、技能、技術棧、學歷、證照、帶隊經歷與聯絡方式
- 對不上的問題交給只看精簡履歷的模型，並附上問題命中的那一段工作內容
- 模型沒有工具、沒有金鑰。問題最多 300 字，10 秒逾時，並限制來源網域與請求頻率
- 逾時、非成功回應或空回答時，改回規則答案

- 同一份履歷: 已知問題
- 只依履歷回答: 其餘問題
- 回到規則答案: 模型失敗時

就是本頁的問答助手，可以直接試。沒有另附評分基準。

---

## Payment and wallet systems

Tech lead, 13+ person team, 2024–2026

Under high concurrency, a resent callback or a network retry can process the same payment twice, and member balances must always match the ledger.

- Payment idempotency, distributed locks, and balance consistency checks, so each transaction posts once
- Compensation and retry for failed transactions, with risk-control audits
- Event-driven architecture on Kafka, Redis, and MongoDB
- Automated reconciliation, plus transaction monitoring and alerts
- Verified with tests for resent callbacks, concurrent requests, and compensation

- No double charges: Idempotency and locks
- Auto reconciliation: Compensation and audits

The company name and business figures are not public. What can be stated is the test coverage: resent callbacks, concurrent requests, and compensation. No live transaction volume is published.

## News crawling and full-text search

AoChuang Technology, senior engineer, 2019–2024

Every new news site needed its own crawler, about 2 days of work, and full-text search took about 2 seconds.

- High-concurrency crawling on Hyperf and Swoole, scheduled and buffered with RabbitMQ
- A configurable crawling framework with rules in MongoDB, so new sites are added by configuration
- Elasticsearch full-text search, kept in sync with MySQL in real time through Canal
- Multi-level caching with Redis and APCu to take load off the database

- ~2 days → under half a day: New site onboarding
- ~2 s → under 300 ms: Search response time
- 100K+: News items processed daily
- Thousands: Items per minute at peak

The search time compares the response before and after full-text search went live. It is not a published percentile, and the query sample size was not kept.

## Resume Q&A assistant

This site · resume tools plus a resume-only model

Visitors ask free-form questions. A model that invents employers, numbers, or dates would make the resume untrustworthy, and a model failure should still leave an answer on the page.

- The page sends known questions to resume tools: company, skill, stack, education, certifications, leadership, and contact
- Other questions go to a model that sees a compact resume plus the bullets of the one job the question matches
- The model has no tools and no secrets. Questions stop at 300 characters, the call times out at 10 seconds, and requests are limited by origin and rate
- A timeout, a failed response, or an empty answer falls back to the rule-based reply

- Same resume data: Known questions
- Resume only: Other questions
- Rule fallback: When the model fails

This is the assistant on this page. There is no separate score.