The latest articles on Forem by pizidavi (@pizidavi). https://forem.com/pizidavi https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3758152%2F170a4330-bac4-4b58-ac8b-efaa4222c90b.jpeg https://forem.com/pizidavi en pizidavi Sun, 15 Feb 2026 21:18:42 +0000 https://forem.com/pizidavi/jurnal-a-voice-first-journal-app-with-on-device-whisper-llm-note-36le https://forem.com/pizidavi/jurnal-a-voice-first-journal-app-with-on-device-whisper-llm-note-36le <p><em>This is a submission for the <a href="https://dev.to/challenges/github-2026-01-21">GitHub Copilot CLI Challenge</a></em></p> <h2> What I Built </h2> <p>Jurnal is a simple mobile journaling app that makes it easy to capture "what happened today" using your voice. Instead of starting from a blank page, you record a short voice note, and the app turns it into a clean, structured journal entry you can skim later.</p> <blockquote> <p><code>Jurnal</code> is a <em>Romanian</em> term that translates into English as <em>diary</em> or <em>journal</em>.</p> </blockquote> <p>The core flow is:</p> <ul> <li>Record audio in-app</li> <li>Transcribe locally on-device using a Whisper model </li> <li>Send the transcription to a larger LLM to turn it into a real note, with sections like summary/highlights/intentions in Mardown</li> </ul> <p>What I like about this approach is the combination of privacy-first transcription with the "writing assistant" part happening after: I still speak naturally, and the app handles organizing the content into something readable.</p> <p>Teck stack:</p> <ul> <li>React Native with Expo</li> <li> <code>react-native-executorch</code> to handle local transcription model</li> <li> <code>@tanstack/ai</code> to provide online LLM utility</li> <li>TailwindCSS by Uniwind for styling</li> </ul> <p>Future improvements I'm excited about:</p> <ol> <li>Custom templates for different styles</li> <li>Choosing custom models for transcription and enrichment</li> <li>Manual editing of the final note in-app</li> <li>Smarter search with local <strong>RAG</strong> </li> <li>A background queue for long-running processing</li> </ol> <h2> Demo </h2> <p>🔗 Repo: <a href="https://github.com/pizidavi/jurnal-app" rel="noopener noreferrer">https://github.com/pizidavi/jurnal-app</a></p> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Home</th> <th>Note</th> <th>Recording</th> <th>Settings</th> </tr> </thead> <tbody> <tr> <td><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fcz8ebp3czyt3zzx445jp.jpg" alt="Home screen" width="591" height="1280"></td> <td><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fmxfkui33xnbu950cm9xz.jpg" alt="Recording new note" width="591" height="1280"></td> <td><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fey8t58jk4uyx6vvpu9t5.jpg" alt="Note screen" width="591" height="1280"></td> <td><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Ftzwu5vzs7sykxbqhcrvh.jpg" alt="Setting screen" width="591" height="1280"></td> </tr> </tbody> </table></div> <h2> My Experience with GitHub Copilot CLI </h2> <p>Copilot CLI was most helpful when I treated planning as a first-class step, not a formality. The biggest productivity boost came from writing a tight plan, implementing in small slices, and then letting the AI validate (and correct) itself.</p> <p>To support that workflow, I made the project intentionally strict: ESLint rules are restrictive, TypeScript is in strict mode with other custom strict types, and I pushed the assistant to run <code>lint</code> and <code>typecheck</code> command before considering anything as done. That created fast feedback loops and prevents subtle regressions from piling up.</p> <p>I also built a custom <strong>Orchestrator</strong> agent that follows a <code>plan → implement → review</code> cycle and repeats until the plan is completed. It was really fun to see the agents correct themselves when they missed a step or introduced a bug.</p> <p>I'm extremely positive about the future of Copilot CLI: the UI is good and the interaction model fits real development.<br> One thing I'd love to see next is a persistent <strong>allowlist for safe commands</strong> (like lint/typecheck/format), so that I don't have to approve them every time.</p> <p>Have fun!</p> devchallenge githubchallenge cli githubcopilot