He Had Never Written a Line of Swift. Then He Built a Better Duolingo.

William Imoh moved to Norway, spent 1,000 days on Duolingo, and still could not hold a conversation in Norwegian. So he built his own app. He had a chemical engineering degree, four years as a product manager, and zero experience writing Swift, the language used for native iOS apps. Using OpenAI's Codex and GPT-Live, he shipped Mural, a real-time voice conversation app, for both iOS and Android.

That is the story. Here is why it matters: the gap between "I have an idea" and "I have a working app" just got meaningfully smaller for anyone who can think clearly about design and articulate what they want.

What Codex Actually Did for Him

Codex is OpenAI's AI coding agent, available at chatgpt.com/codex. Think of it less like autocomplete and more like a junior developer you can direct entirely in plain language. Imoh used it to write the iOS and Android code for Mural from scratch, guiding it with prompts rather than writing Swift himself.

The other key ingredient was GPT-Live, OpenAI's real-time voice API. This is the engine that lets Mural hold a live spoken conversation with a learner in Norwegian, responding naturally rather than waiting for a full sentence to be typed. Imoh has spoken about latency, the tiny delay between you speaking and the app responding, as a threshold that makes or breaks the experience. Too slow and it feels broken. GPT-Live kept that delay low enough to feel like a real conversation partner.

The result looks nothing like Duolingo. That was intentional. Imoh describes using a specific design prompt focused on visual language and aesthetic direction, and the app came out with a distinct identity. This is not a clone. It is a considered product.

Why a Designer Should Pay Attention

Here is the thesis, plain and simple: design articulation is now a core building skill.

Imoh is not exceptional because he learned to code fast. He is exceptional because he already knew how to describe what he wanted with precision, a skill he built over four years as a product manager and through his work at Hackmamba, a developer marketing company. He knew what a good language-learning experience should feel like. He knew what visual identity he was after. Codex handled the syntax. He handled the thinking.

For UX and product designers, that split is very good news. The part of building that has always been hardest for non-engineers, writing correct code in unfamiliar languages, is now something a well-directed AI agent can carry. The part that remains hard is the part designers are already trained to do: knowing what to build, for whom, and what it should feel like to use.

This is exactly the kind of project that fits naturally alongside other vibe-coded projects in the making, where the idea and the experience come first and the technology follows the brief.

A Practical Look at How He Ran the Whole Thing

Imoh runs Hackmamba, a team of roughly 25 people, largely solo. He uses Codex Projects, a feature within Codex that lets you organise your work by context so the AI stays focused on the right codebase or task at a time. For the rest of his company operations, he built what he calls a "fellowship" of Grok Bot agents, each one assigned a different solo-founder role and named after a Lord of the Rings character. Each agent has access to relevant context, though he is deliberate about what he keeps off-limits.

For a designer starting to build, there are two practical takeaways from his setup. First, Codex Projects can act like a focused workspace. Rather than starting every session from scratch and re-explaining your app to the AI, a project gives it persistent context about what you are building. Second, the agent approach to running a business suggests that AI tools are not just for writing code. They can hold context for decisions, drafts, and operations across your whole workflow.

The Honest Limits

Imoh is clear about what he still has to do himself, though the source does not detail every task. He has also said the fear that keeps him up is not being replaced by AI. It is becoming too dependent on it. That is a genuinely useful distinction. The risk is not that these tools take your job. The risk is that you stop developing judgment because the tool always answers first.

Mural is a real app, with a public GitHub repository and a live site. It was built by someone who, by his own account, had never written the native code it runs on. That is worth sitting with. Not because it means anyone can build anything overnight, but because it shows that knowing your user, knowing your design intent, and being able to describe both with clarity is now most of the job.

The tools are ready. The question is whether you have something worth building.