Building products with AI.And helping your business go global.
I’m Garfield, a full-stack engineer and independent developer. Since 2022 I’ve helped build Caixuan, Gezhe and DeckFlow, all tools for Office and iWork documents like PowerPoint and Keynote. Now I bring that engineering experience to working with AI: I set the goals, give it the full context and design the architecture; AI agents write and deploy the code in parallel; then I review the result against the standard I set.
I use the same approach for people without a technical background who want to sell overseas. Customers abroad should be able to find you online, understand what you offer and buy directly.
How I divide the work
My main development tool. One session coordinates architecture and tasks while other agents implement them. I focus on goals and design decisions.
Handles browser tasks, account setup, UI testing and some code review. It also takes on coding work.
Voice input for context, requirements and feedback. It lets me keep a project moving from my phone, as I did with OfficeLint.
Now2026-11
I’m working full-time on DeckFlow. It started with slides and is becoming a document workspace: upload any document to have it parsed and extracted, then find it fast and share it whenever you need to. Office files can be edited directly. You can also generate new documents, refresh old ones or translate them into other languages.
Both people and AI can use it. Complex jobs run through complete workflows; simple ones take two or three steps. AI agents call the same capabilities directly through a CLI skill.
The rest of my time goes to helping people with no technical background, who rarely use AI, move their business online. Most recently, a Mandarin teacher got a site of their own, Talkdry: students abroad find the teacher online, pay tuition directly and take lessons over Zoom.
I also help people running overseas businesses plan their finances and set up the flow of money: collecting payments → paying out → investing for growth.
Selected work
DeckFlowAI document workspace
It began with document parsing and slide generation, translation and redesign, and is growing into a document workspace: upload, search and share any kind of document; edit Office files directly; generate, refresh and translate. People use it through workflows and small tools; AI agents call it through a CLI skill.
More than 220,000 users as of November 2026.
What I did
- I’ve worked on it since the first release, responsible for the backend and platform APIs.
- Document parsing. I extract content and structure from Office files and PDFs. Translation, redesign, search and the agent tools are all built on this layer.
- AI workflows. I design the context building and DAG-based orchestration that break complex document tasks into steps that can be combined and reused.
OfficeLint
I shipped a document tool while hiking, directing AI agents by voice from my phone. OfficeLint checks, edits and redacts Word, Excel and PowerPoint files. The experiment was to find out how much of development I could delegate while away from a computer.
Six Claude Code sessions shared the development work. I spoke to the architecture session, which assigned tasks and coordinated deployment. Codex used the browser to configure payments, authentication and cloud services. The product is live, and I recorded the process.
MacroSora
A patient support service for international visitors seeking care in China. I started with technology and search acquisition: a site in five languages covering services, quotes, bookings and payments, alongside coordination with hospitals and service partners.
Two weeks after launch, our first American customer found us on Google. After a week of conversations, he decided to travel to China and subsequently had surgery, with our support throughout. That first delivery tested the full process from search to trust to in-person service. The business is still at an early stage.
Talkdry
An independent course site for a Mandarin teacher serving students abroad. He wanted his own channel for enrollment and payments. I connected course information, Stripe payments and Zoom lessons so prospective students arriving from social media could learn about the course and pay in one place.
Built in a few hours by directing AI remotely through voice. Two students have paid through the site’s Stripe checkout, verifying the payment flow with real orders. The next step is bringing in more students.
Gezhe PPT
Turns a topic or source material into a slide deck. The team shipped the first version in two weeks, placing text from GPT-3.5, then the most advanced model available, into prepared templates to test demand.
600,000+ registered accounts as of November 2026.
My role: Backend APIs, the AI SDK, integration and deployment.
Caixuan PPT
A slide collaboration and sharing platform built from scratch, and an early product from the team behind Gezhe and DeckFlow.
My role: Backend development, integration, deployment and operations. I set up CI/CD to automate builds and deployments after code was pushed.
How I check the work
- 01
Set the goal first, so there’s something to check the result against.
On engineering work, we first discuss why we’re building it, what it should become and how it will be judged. For PDF parsing, I defined the acceptance criteria, built a test set and set up an evaluation loop before letting AI drive the implementation.
- 02
Have AI check its work against the original brief.
I keep asking: “What was the task I gave you? What were the requirements? Which ones have you finished? Are you really done?” AI then compares what it produced with the original goals, item by item, and finds what it missed.
- 03
Keep a person checking the reasoning and the build at key points.
For architecture, I have AI draw the design as a diagram, then we review it together: what is overly complex, and is there a better approach? Where my own thinking isn’t clear yet, I turn it around and have AI give me ten questions that expose what I don’t know or haven’t considered. That round of questions brings missing requirements and constraints into the design before work continues. AI helps propose solutions and find blind spots; I need to understand the trade-offs and make the call.
- 04
Test AI’s judgment against what actually happens next.
With MacroSora, I wanted to see whether customers abroad could find a patient support service through SEO and GEO, and would want to talk further. Two weeks after the site launched, our first American customer found us on Google; after a week of conversations, the customer came to China for treatment. That delivery gave us the first complete record of acquisition, trust-building and in-person service. Whether the business lasts will take more customers to prove, and more iteration.
- 05
Turn corrections into lasting practice, so the same mistakes happen less.
When AI keeps making the same kind of mistake, I write it up as a skill. Project facts go into context I keep up to date; reusable practices and checks become rules. What one review uncovers changes how AI carries out the next task.
Outside work
I enjoy hiking and talking with people in different industries. I also keep experimenting with AI in my everyday life and work.
Contact
Have a product in mind, a workflow to improve, or want to compare notes on working with AI? Find me on WeChat: dudu_22, or email [email protected]. My code is on GitHub.