← Jonáš Světlík
06 - JARVIS · PERSONAL ASSISTANT
CSEN
online 24/7 · built from scratch

Jarvis.
My personal
AI assistant.

I built my own Jarvis - entirely from scratch, in my own code. He runs non-stop on a Raspberry Pi at my place: I talk to him out loud, he writes me morning briefings, he knows my health data from a Garmin watch, manages my servers and remembers what we have been working on.

I want my own assistant → AI agents ↗

# jarvis.status · STARK-style HUD

BrainClaude · auto / deep
Voice● call mode · live voice
HealthGarmin · sleep · HRV
Servers2 machines watched

What he can do

An assistant that genuinely knows me.

Speaks out loud

A voice call like a phone call - a hands-free loop, a neural voice and a WebGL visualiser that pulses with the speech. Tap to interrupt him.

Morning briefing

Every morning at 7:30: how I slept, whether to train, the weather with a suggestion on what to wear, my reminders and the state of my servers.

Knows my health

Data from a Garmin watch - sleep, heart rate, HRV, Body Battery. "Should I train today?" and he answers from real numbers.

Manages servers

Over SSH he watches both the Raspberry Pi and the production server. Read-only commands he runs himself; every change he puts to me for approval first.

With me everywhere

Telegram - I write or send a voice note, he answers in text and in voice. On the iPhone as an app, working from anywhere.

He remembers

My own notes, a knowledge base from my Obsidian vault and an automatic journal - he writes down how I am doing and what I am planning.

Under the hood

Built by hand, not out of a box.

Brain & knowledge

How he thinks

  • Claude as the brain - a fast model for everyday questions, deep thinking for the hard ones; switchable on the fly
  • A knowledge base from Obsidian - he knows my notes and my projects and answers in context
  • An agent with tools - he runs a command himself, sets a reminder, searches the web
  • Journal and memory - after every conversation he writes down what is new, in the background
  • Voice - voice messages transcribed locally (Whisper), answers in neural TTS
Hardware & hosting

Where and how he runs

  • Raspberry Pi 5 at my place - running 24/7 in Docker, drawing a few watts
  • Web + call + Telegram + iPhone - one Jarvis, four ways to reach him
  • HUD dashboard - the state of both machines (CPU, RAM, disk, containers) in real time
  • Safety catches - destructive commands blocked, changes only with my confirmation
  • My data stays at home - the journal, the health data and the notes all sit on my own hardware

A typical morning

07:30 · "Good morning, sir."

Before I am up, Jarvis pulls fresh data off the watch, checks the weather and the state of my servers - and sends me a briefing on Telegram: how I slept, whether to train hard today or take it easy, what to wear, what is on my plate. If the watch has not synced yet, he waits for them and only sends the briefing once it has real numbers.

Garmin sync→ sleep + recovery→ weather→ reminders→ a briefing on Telegram

Stack

What it runs on.

PythonFastAPIClaudetool-use agent DockerRaspberry Pi 5Whisperneural TTS Three.js / WebGLTelegram APISQLitePWA

Jarvis is a non-commercial personal project - he is not for sale; I show him as an example of what I build and what I learn on. Alongside him I am also experimenting with a desktop voice assistant built on the open-source project Mark-XLVII (by FatihMakes, licensed CC BY-NC 4.0), which I extended with features of my own - but everything you see on this page is my own code from scratch.

Want a Jarvis of your own?

Just as I built mine, I can build an assistant to fit you - from scratch, in my own code. Voice, automation, your data and your channels.

Get in touch →