People make the calls.AI does the build.
Kaiwu Tech is one person, Kai Wu, plus a team of AI agents. Kai Wu works inside your workflow to find the problem, decides how to fix it, and signs off on the result. The agents write the code, run the tests, and prepare the handover documents.
Find the problem first, then decide whether AI fits
Writing code and building tools is no longer the bottleneck. What companies are missing is someone who sees which part of the workflow eats the time and which numbers nobody can reconcile.
So the first step is a 30-minute intro call. You describe how the work gets done today, and Kai Wu judges whether it is worth a closer look and where to start. Thirty minutes won't find the real problem; that happens next, when we look at the actual workflow. Then we write down what “fixed” means, and only then start the work.
The fix depends on the problem: sometimes a process change or cleaner data, and only sometimes code or AI. Old and new run side by side, and we switch over only when the numbers match. You get the operating manual, how each number is calculated, and the source code.
Services and how we workWho does what
Kai Wu handles
- Finding the problem inside the workflow
- Business judgment and trade-offs
- Agreeing on goals with the client
- Acceptance testing and quality control
AI agents handle
- Writing the code
- Automated tests
- Documentation and handover materials
We use this method on ourselves first
From May to July 2026, Kai Wu and the AI agents ran 12 projects in parallel and made 2,000+ code commits. A single project, the weekly report automation, has about 700 automated tests guarding the full pipeline.
The work covered report automation, intranet data scraping, data dashboards, a website migration and an ad attribution system. None of it was handed to an outsourced team you never see. It runs on a structured process (plan, execute, verify), and every project keeps its own memory file.
The method is written up as flightwake, and client projects use the same one.
Source: the 12 projects and 2,000+ commits come from git commit statistics for May 1 to July 4, 2026. The test count is from the weekly report case study.
Three things we hold to
Replace safely, not recklessly
Fixing a workflow isn't just switching off the old one and turning on something new. A parallel run, a rollback plan and handover documentation are included. Before an old process is retired, the numbers have to show that the new one can be trusted.
Solid data first, then AI
If the data can't be trusted, AI only reaches the wrong conclusions faster. We integrate first, then standardize and confirm the data is reliable, and only then bring in AI. We don't skip steps in that order.
Legacy systems and intranet data can be fixed too
We have automated legacy systems that could only be viewed over a VPN. A system you can't change, or data locked behind an intranet, is not where improvement has to stop.
Who we work with
Companies of 10–100 people with process-heavy, repetitive work that are willing to pay for efficiency, especially professional services: accounting and bookkeeping, legal and land-registration services, consulting, HR, insurance, and medical clinics. If someone at your company spends several hours every week on a task whose steps are the same every time, that is a good place to start the review.
Tools we use
- Supabase
- Odoo
- n8n (self-hosted)
- Docker
- Next.js / Vercel
- Playwright
- Claude / AI agents
Tools serve the workflow, not the other way around. We choose them for stability and reliability, not to show off technology.