Haru, Korean for "a day", is a Korean-language personal AI assistant in development for natural voice conversations, information retrieval, desktop actions and study. You talk to it, and it looks things up, opens apps and sites on your PC, keeps your schedule and reminders, and helps you review lectures.
Under the hood
An unedited screen recording. Haru answers by voice, briefs the news, then opens a chart and a study playlist on request.
Haru is a general personal assistant, not a study-only app. Learning support is one area among several. Each feature shows its current status.
Wake it with a wake word or a double clap, or switch on always-listening and just talk. Speech is transcribed on the PC with faster-whisper, and replies are read aloud in a voice generated on the user's own GPU. You can interrupt it mid-sentence. Follow-up conversation is still being improved.
Web search, weather and air quality, currency, Wikipedia summaries, holidays and tech news, all through real free APIs. Live facts come from tool results, not guesses.
It launches registered apps, sites and games, opens sites in your logged-in Chrome profile, plays YouTube or Spotify music, and handles do-not-disturb, in-app notes and window layout. Common commands are handled instantly by local rules.
Save an event by voice and it reminds you 10 minutes before and at the start. It reads your phone calendar (ICS) for briefings and keeps memory on this PC. Writing to external calendars and server-side long-term memory don't exist yet.
Chart wall, economic calendar and news, order-flow and quant analysis, Bitcoin on-chain data and pension-fund disclosures. Price and deadline alerts are checked by code, not a model. No performance claims.
Class mode (silent marking and a sourced review) is implemented on the PC. A study helper built as a Claude artifact makes notes, practice questions, flashcards, mind maps and PDFs. So far it has only been tested with stub responses.
A separate page with recording, local recovery, Groq Whisper transcription and timetable recognition. It isn't deployed yet and hasn't been verified on a real iPad or with real model responses.
Per-user sign-in comes first. After that comes connecting email, calendar and Notion accounts the user approves. Today the prototype is protected by a single shared password, so sign-in has to change before any personal accounts are linked.
Unedited screenshots from the developer's Windows PC. "JARVIS" on screen is the codename used during development; the product name is Haru.





Fast local rules, Claude's judgment and other models each have their own part. Here is the split.
Open questions go to Claude Code, logged in on the PC and allowed to read and web-search only. If Claude Code is missing or fails, a free Groq model answers instead.
Spoken feature requests go to the Claude Code CLI, which edits a copy of the app. Syntax checks, test runs and user confirmation come before it's applied, with automatic rollback.
The web server has a Claude API path with tool use and web search. The current deployment has no Claude key set, so Groq answers today.
A Claude artifact flow analyses lecture material into notes, practice questions, flashcards, mind maps and PDFs.
What Claude doesn't do: transcription is Whisper (faster-whisper on the PC, Groq on the web and recording page). Request types (analysis, research, action, development), common commands and alerts are code rules. The agent nodes on screen are visual only.
A university student and solo developer in South Korea, building Haru before registering a business. No users, revenue or funding yet.
Product feedback, beta testing and partnerships are all welcome. The founder replies personally.
minjun@heyharu.kr