iOS/macOS developer in Taipei. 9+ years shipping production apps. Right now I'm building my own agent from scratch — and publishing a build log every single day.
Marvin Builds AI is my YouTube channel about real AI-agent workflows. I hand agents actual tasks and watch how they read requirements, run code, fix mistakes, deploy, and verify results.
Not demo reels — the point is the control flow behind the demo: how coding agents and automation become workflows you can verify, reuse, and actually keep control of.
Not another coding agent — those already exist and they're good. Mine solves a different problem.
It's built for whoever does the acceptance testing. In most teams that's PM and QA, and they're the ones actually blocked: they can't build the latest branch, and they can't easily prove a bug is really fixed in this build. Codex and Claude are available to them, but someone still has to set up the repo, environment, credentials and workflow first. The tools exist; the barrier doesn't go away.
So the agent wraps what developers have already set up, and hands it over as something a non-developer can just use:
"Build me the latest version." "Confirm whether this bug is fixed." "Verify this feature against the ticket."
First capability — specification acceptance testing. Start from a Jira ticket or plain-language acceptance criteria → switch to the target branch → prepare the environment → build for iOS Simulator or Android Emulator → drive the app → capture screenshots as evidence → generate HTML and PDF reports → deliver to a chat channel.
Signature Abilities — capabilities unique to this agent, each with its own GUI page so they're usable directly instead of staying buried behind a prompt. The first is spec-diff: give it an old spec and a new spec, get back a structured, human-readable report of what changed, what was added, and what was removed. It generalized further than planned — contracts, architecture designs, even trading rules.
Currently in Alpha.
📓 Agent Build Log — one episode a day, bilingual (繁體中文 / English).
| Project | What it is |
|---|---|
| Earnings Rosetta | U.S. earnings-call platform — verbatim transcripts with AI-generated summaries, so you can grab the highlights fast. iOS & Android |
| Stock Toolbar | macOS menu-bar app: type a ticker, get the Taiwan stock price — without leaving what you're doing. |
| King of Song Quiz | Song-guessing game. Five songs a round, faster correct guess scores higher. My first shipped game. |
| Backwards Clock | iOS tribute to Grace Hopper's counter-clockwise office clock. (Off the App Store) |
Platforms — iOS · macOS · Android Languages & frameworks — Swift · SwiftUI · Flutter / Dart Backend — Cloudflare Workers · D1 · KV · Python (Django / FastAPI) Practice — clean architecture · unit testing · CI/CD · serverless
Bilingual posts (繁體中文 / English) on marvinswift.com: Agent & Programming · Swift · Unit testing · Videos