Find the problem first.Then decide where AI fits.
AI can already write the code and build the tools. The bottleneck is seeing which part of the workflow eats the time and which numbers nobody can reconcile. We work inside your workflow to find the problem, then fix it in a way that fits how your team works.
Kaiwu Tech is one person, Kai Wu, plus a team of AI agents. The person finds the problems and makes the calls; the agents do the build.
Or message us on LINE: kaiwu0919- 去除「AI 做的樣子」，定位改為「先找出問題，再談 AI」
- GSC 回查、impeccable audit 修正、英文版(/en)與一篇新文
- 補兩篇 blog（8/14、8/15）＋ 新增 /privacy 隱私權政策頁
- 清掉三條掛著的品牌待決:主張句穩重化、favicon 容器版、og 主圖換 V5 標記
15 flight records since 2026-07-18, kept in Traditional Chinese. Decisions go into DECISIONS, pitfalls into TRAPS. Client projects use the same system.
The method: flightwakeProblems we're usually called in to look at
Most people come to us with one of these. We first work out where things get stuck and why, then decide how to fix it. Sometimes that means changing the process, sometimes cleaning up the data, and only sometimes writing code or adding AI.
- Someone spends hours every week building the same reports
- Your data is locked in legacy systems or an internal network, and head office has no live view
- You spend on ads, but can't tell which ones lead to sales
- Each department's numbers are spread across a dozen Excel files, and the versions never match
- Company know-how lives in senior staff's heads, and one question from a new hire interrupts three people
- You can't update your website, and the source code isn't in your hands
Case study・Dental group in northern Taiwan · 35 branches
A 31-KPI weekly report only certain staff could build, with numbers nobody could trace
The marketing team manages 13 Meta ad accounts. Before every weekly operations meeting, someone had to build 6 sub-reports covering 31 KPIs by hand. The problem wasn't only the time: every step was scripts run by hand and copy and paste, only certain staff knew how to do it, and when numbers didn't match nobody could say which fields they came from.
The problems we found
- Every week someone ran scripts by hand, copied spreadsheets, exported CSVs and pasted them back. Any step could go wrong
- 6 sub-reports and 31 KPIs, with data from the ad platforms, customer service logs and doctor schedules
- When numbers didn't match, nobody could say which fields and rules a given KPI was calculated from
What we changed
- 1Data first: ad APIs write straight into the database, all 13 accounts sync in one run, and re-running it never corrupts the data
- 2Safe cutover: old and new systems ran side by side for 4 weeks with a cell-by-cell comparison (tolerance ±1). The old process was retired only after the numbers matched
- 3A 100-item acceptance checklist signed off item by item, with a full operations manual and a tool for backfilling historical data
- 4One-click weekly report: 7 automated steps handle the import and all 6 sub-reports, and every KPI has traceable data lineage
Results
- Weekly manual reporting → one click, no longer dependent on specific staff
- All 31 KPIs have documented formulas, and every number can be traced to its source
- About 700 automated tests continuously protect the whole pipeline
The hardest part of automation isn't writing the code. It's getting the owner confident enough to switch off the old process. Four weeks of parallel runs and cell-by-cell comparison earned that trust.
How working together goes
- 1
A 30-minute intro call
Free. You describe how the work gets done today, and Kai Wu judges whether it's worth looking into and where to start. Thirty minutes won't find the real problem, so the next step is the one that matters.
- 2
We look at the actual workflow
This is where the real problem-finding happens. We watch the work run once: which tools you use, where the data comes from, which step only one person knows how to do. If you have a requirements document, bring it.
- 3
We agree on what “fixed” means
Before changing anything, we write down the acceptance checklist and which numbers to compare while old and new run side by side.
- 4
We fix it, then run old and new in parallel
The fix is whatever fits how your team works: a process change, cleaner data, or code and AI where they help. We switch over only when the numbers match.
- 5
Handover
You get the operating manual, how each number is calculated, and the source code. You own all of it.