PersonalOngoingFinished2025

Haru

Built entirely by hand, Haru is a massive productivity ecosystem that has expanded to function almost like a specialized study operating system. The platform helps users research and learn topics faster by rewarding focused work and punishing distractions. It merges offline privacy and real-time online updates by running local AI models (via llama.cpp and fastText) alongside web search tools. Key features include localized RAG across user notes, cognitive minigames, Todoist/ActivityWatch integrations to track computer usage, built-in Excalidraw, advanced calculators, Pomodoro timers, and speech-to-text/text-to-speech engines that generate clean, Obsidian-compatible markdown notes. Active development was completed to focus on smaller, targeted sub-projects after extracting immense learning value from this codebase.

Haru

About This Project

Designing an all-encompassing productivity ecosystem that tracks workspace state, compiles markdown summaries, runs local AI inference, and incentivizes focus while eliminating distractions.

Engineered a custom desktop environment using SolidJS and Tauri. Orchestrated llama.cpp for offline RAG, integrated OS-level window and activity tracking via ActivityWatch/Todoist, and implemented speech transcription pipelines to output markdown compatible with Obsidian.

A highly customized 'productivity OS' that unifies notes, tasks, calculators, and mindmaps into a single window. The project served as a massive systems design playground before being completed in favor of smaller modular tools.

Role

AI & Full-stack Developer

Year

2025

Status

Finished

Type

Personal

Technology Stack

TauriRustSolidJSPythonWSLFastTextLlama.cppFFmpegObsidianTodoistActivityWatch

Project Story

The Challenge

Designing an all-encompassing productivity ecosystem that tracks workspace state, compiles markdown summaries, runs local AI inference, and incentivizes focus while eliminating distractions.

The Approach

Engineered a custom desktop environment using SolidJS and Tauri. Orchestrated llama.cpp for offline RAG, integrated OS-level window and activity tracking via ActivityWatch/Todoist, and implemented speech transcription pipelines to output markdown compatible with Obsidian.

The Outcome

A highly customized 'productivity OS' that unifies notes, tasks, calculators, and mindmaps into a single window. The project served as a massive systems design playground before being completed in favor of smaller modular tools.

Insights & Takeaways

Highlights

  • Functions as a comprehensive, hand-built 'productivity operating system' housing tools like Excalidraw, Pomodoro, and cognitive minigames.
  • Features OS-level activity tracking via Todoist and ActivityWatch integrations to reward focus and warn of distractions.
  • Generates Obsidian-ready markdown notes through integrated Speech-to-Text (STT) and Text-to-Speech (TTS) pipelines for audio files and video playlists.

Challenges

  • Orchestrating multi-layered local binaries (llama.cpp, fastText) and speech model pipes under a single SolidJS UI without degrading responsiveness.
  • Connecting third-party productivity APIs (Todoist, ActivityWatch) to local database records dynamically.

Lessons Learned

  • Building a massive, monolithic 'personal OS' provides unmatched architectural training, but also proves that small, single-purpose utilities are often more maintainable.

Related Work