1,000 days of Duolingo didn't work. So he built his own app with Codex.

He moved to Norway expecting to get by in English, but found it hard to practice Norwegian because people would switch to English when his Norwegian was imperfect. His goal became finding a conversational partner that wouldn’t judge him and would correct him in real time.

He framed Mural Chat as a premium-feeling native app and chose Swift for iOS and Kotlin for Android despite never having written Swift. His point was that with AI agents, the limiting factor is no longer the developer’s existing language knowledge, but picking the best tech for the experience.

He started with a long, highly specific prompt focused on aesthetics and interaction design (a warm, autumn-like palette, minimalist UI, a pulsating orb, and only necessary elements). He iterated after the first output didn’t match his taste (e.g., rejecting dark mode) and refined the design direction with clearer constraints and references.

In the demo, the voice agent corrected grammar in-line (e.g., correcting “en møte” to “et møte”) and refused to switch to English, repeatedly nudging him to respond properly in Norwegian. William described this as the core value: relentless, nonjudgmental correction until you say it right.

He said running the real-time voice experience is expensive, estimating roughly $3–$5 for 30 minutes of conversation, and described funding it with a personal OpenAI budget while keeping the project open source. He framed it as a fun project rather than a business he wanted to maintain under user expectations.