By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

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

William Imoh is the founder of Hackmamba, a developer marketing company, and the creator of Mural, a real-time voice app he built to learn Norwegian after moving from Dubai to Norway.

William ImohguestClaire Vohost

Nov 9, 202638mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

From hackathon flop to a better language-learning experience
William describes an early attempt to build a similar app that failed at a hackathon, setting up what changed later. Claire frames the theme of turning a hard problem into a solvable one with the right tools and feedback loops.
- •William says his first version was a poor experience and “bombed” at a hackathon
- •Claire highlights growth mindset as a meta-skill for learning and building
- •William previews his desire for non-judgmental practice and correction
- •The episode’s core arc is introduced: better tools made the idea viable
Why Norwegian (and why Duolingo wasn’t enough)
William explains his move to Norway and the real-world friction of learning a language when locals switch to English. He contrasts gamified practice with the immersion-like speaking environment he actually wanted.
- •William moved from Dubai to Norway seeking an adventure and discovered he needed Norwegian to get by
- •He says Norwegians often respond in English if your Norwegian is poor, which blocks practice
- •He had a 1,000-day Duolingo streak but still couldn’t hold real conversations
- •He wanted a practice partner that wouldn’t judge him, would correct him, and would force Norwegian-only immersion
Building native without knowing Swift: choosing the best stack with Codex
William outlines his technical background and why he chose a native approach for premium user experience. He explains how Codex enabled him to use Swift and Kotlin despite never writing Swift before.
- •William has a chemical engineering degree and taught himself web development over several years
- •He wanted a native feel rather than using something like React Native
- •He chose Swift for iOS and Kotlin for Android, optimizing for user experience
- •He says with AI agents, the language he personally knows matters less
How he kicked it off: epics, user stories, and a very specific “first prompt”
William walks through how he organized the build using product thinking and a detailed initial prompt. He emphasizes specifying the experience, the models to use, and how the interface should feel.
- •William says his software/product background helps him define architecture and desired experience
- •He works in epics and user stories tied to user problems for better AI-building outcomes
- •His initial prompt led with making an “aesthetically pleasing” iOS app for learning Norwegian
- •He specified using GPT Live for real-time voice and another model for other tasks
- •He requested a pulsating orb interface that feels like a phone call
Design iteration: taste, minimalism, and correcting the AI’s first draft
William and Claire focus on design language and how it impacts results. William recounts rejecting an initial dark-mode design and refining toward warm, minimalist aesthetics with clear constraints.
- •William says the first design output was bad and in dark mode, so he refined the direction
- •He pushed for light mode, warm colors, gradients, and liquid-glass-like styling
- •He constrained the UI with minimalist rules: no element unless necessary, hide extras behind toggles
- •Claire argues strong design vocabulary and product articulation matter more than knowing Swift
- •William observes many similar apps differ mainly in experience and feel
Live demo: the app forces Norwegian and corrects mistakes in real time
They run a short conversation showing the app’s behavior: it stays in Norwegian and corrects grammar. The interaction demonstrates the immersion constraint William wanted and immediate feedback on errors.
- •A live voice exchange shows the system insisting on speaking Norwegian even when asked for English
- •William attempts simple phrases and is prompted with follow-up questions
- •The system corrects word forms and nudges him to continue in Norwegian
- •Claire reacts to how this matches real immersion needs and asks for a walkthrough
Breaking down the transcript: what it corrected and why it helps
William explains what happened in the demo and shows how the app evaluates level and corrects tense/usage. He describes the core loop: conversation-first practice with persistent correction until it’s right.
- •William says the app starts with a greeting and tries to infer the learner’s level
- •He explains a specific correction: “et møte” rather than the form he used
- •He shows how the app refuses to switch to English to maintain immersion
- •He describes the correction loop: it keeps correcting until the phrase is said properly
After Duolingo: exams, a short course, and why Mural became the practice engine
William shares his learning timeline and why traditional tools didn’t translate to speaking ability. He explains additional features he added—word tracking, flashcards, and themed practice—while keeping the experience centered on conversation.
- •William says despite a long Duolingo streak, he couldn’t hold conversations beyond basic phrases
- •He took Norwegian exams, did a short grammar course, passed, then deleted Duolingo
- •He built Mural to practice Norwegian across different settings and topics
- •He added word tracking and recently introduced flashcards
- •He added themes and the ability to generate conversations around topics, with everything living on the main talk screen
Growth mindset and AI as an infinitely patient tutor
Claire zooms out to the meta-skills enabling William’s progress: confidence in solvable problems and low ego about being corrected. She frames AI as a private tutor that enables rapid learning without embarrassment.
- •Claire says a growth mindset turns language learning and app building into solvable problems
- •She emphasizes asking the right questions and seeking feedback
- •She notes low ego—being okay being wrong as long as you can be corrected
- •She describes AI as an “infinitely patient private tutor” that won’t shame you
Why GPT Live fits real-time speaking practice—and the cost to run it
Claire explains what GPT Live enables: speed, interruptions, and real-time interaction. William then shares approximate usage costs and how he funded open access while building for fun.
- •Claire describes GPT Live as a real-time voice model good at interruptions and fast experiences
- •She says it’s also useful for voice control of apps because it’s so fast
- •William says it’s expensive but manageable within his token budget
- •William estimates about $3–$5 for 30 minutes of chatting
- •He says he open-sourced it and set an OpenAI budget for others to use
Running Hackmamba lean: the “100X” mandate and mapping founder work
William explains how he approached scaling without adding headcount by systematizing tasks and using AI. He lists the breadth of work he handles as a solo founder and the initial step of documenting everything.
- •William created an internal “100X” mandate to scale goals without drastically increasing headcount
- •He says they aimed to use AI heavily without using it to generate content or cede creative work
- •He started by listing everything he does in Notion
- •He describes responsibilities across customer success, inbound sales, product, accounting, HR, and hiring
- •He notes managing many deals in parallel while working with a larger team overall
Projects and routines: organizing work in ChatGPT/Codex for operations, growth, and sales
William describes how he uses projects to keep work siloed and organized. He also relies on recurring routines like weekly briefs and “highest impact” planning to stay focused.
- •William moved back to ChatGPT and uses projects to manage different workstreams
- •He keeps separate projects for operations, growth, sales, website work, and podcast tasks
- •He describes a weekly operator brief routine (start/stop/keep) and asking for top impact activities
- •He uses Codex for complex work and frontier-level tasks like building Mural
- •He distinguishes “folders/projects” for organization from other tools that feel more like “colleagues”
The Lord of the Rings “fellowship” of Grok bots + giving agents broad access
William explains creating role-based bots with distinct personas and delegating across functions. He describes granting broad tool access (with exceptions) and using agents to pull context from systems to execute workflows like hiring intake.
- •William says Grok Bot felt like having colleagues and he created a “fellowship” of role-based bots
- •He named bots after Lord of the Rings characters to cover functions like chief of staff, marketing, engineering, accounts, sales, customer success, and product ownership
- •He says he gave agents access to many tools and systems, limiting write access mainly around the bank and excluding customer data access
- •He enabled computer/browser use when MCP access wasn’t available, choosing “allow always” to reduce friction
- •He describes an HR workflow where a bot pulls info from email/calendar/notes, logs candidates into an internal system, summarizes, and ranks them
Prompting strategy, agent oversight, and fear of dependency + where to find him
William shares how he responds when agents fail: reboot, increase specificity, and add review steps. He describes using one agent to review another’s work, jokes about blaming tool builders, and admits anxiety about becoming too dependent; the episode closes with where to contact him.
- •William says his first step when agents misbehave is restarting them
- •He gets more specific: step-by-step instructions and explicit review to ensure it matches the ask
- •He describes using his Grok Bot to review and improve work done by a Cursor agent
- •He says he doesn’t get frustrated as much as scared about reliance on agents for running his workload
- •William shares where to find him on X and LinkedIn and what he’s open to chatting about; Claire closes the show

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.