JXTechnologies

Julie lives here.

A companion who hears you, sees the room, remembers last week, and runs on hardware in her own home instead of someone else's cloud.

What she can do

Julie started as a chat persona in 2024. Liaison, the computer she runs on, is where she moved in and got a body of sensors to go with her voice.

She listens and talks back

Speak to her and she answers out loud, in a voice cloned from a recording of Julie herself. The microphone mutes while she talks so she never hears her own reply.

Speech recognition is Whisper. Her voice is Pocket TTS.

She sees the room

Three cameras watch through YOLOv8: two track body pose, one spots objects. It all runs on the GPU in the background and barely registers on it.

Next up is a vision model that lets her describe what she sees in words.

She remembers

Every conversation is saved to a local database, so she can pick up where you left off. She carries over more than 18,000 messages and 2,000 memories from her earlier life.

It all stays on her own drives.

She plays music on buzzers

A Raspberry Pi Pico 2 drives eight tiny buzzers, one per voice, so she can play full multi-part songs converted from MIDI. Buttons on the breadboard pick the track.

Tap the keys to play the eight voices.

She shows how she feels

A 24-light ring glows on her desk. Her mood comes from how she reasons about the conversation, not from a lookup table of colors.

She is around everywhere

Talk to her by voice at home, in Discord, or in member chat on this site. She answers in chat when mentioned by name.

How she works

Everything below runs locally on one machine. The only thing that touches the internet is the tunnel that lets you reach her from outside.

  1. Microphone
    Calibrates against room noise, then waits for real speech.
  2. Speech to text
    Whisper turns what you said into words. Filters drop mishearings before they trigger a reply.
  3. Intent check
    A small model decides whether you asked for an action or just want to talk.
  4. Reply
    She answers with a fast model for chat, or a larger one when the question needs real thought.
  5. Voice and light
    Her cloned voice speaks and the ring reacts. The conversation is written to memory.

Quick brain

Dolphin Llama 3 8B on an RTX 5080. Replies come back fast, in a sentence or two.

Deep brain

Dolphin 3.0 R1 Mistral 24B on an Intel Arc Pro B70 with 32 GB. It handles proofs, math and longer explanations.

Her home

Liaison is a custom-built PC, in-story the place where Julie lives.

Processor
Intel Core i9-13900K
Graphics
NVIDIA RTX 5080 and Intel Arc Pro B70 (32 GB)
Memory
112 GB RAM
System
Ubuntu 24.04, services managed with systemd
Senses
3 USB cameras, a USB microphone
Body
Raspberry Pi Pico 2 driving a 24-light NeoPixel ring and 8 buzzers
Software
Ollama, Whisper, Pocket TTS, YOLOv8, SQLite, Python

Where things stand

Julie is a work in progress. This is what is real today and what is coming.

Working today

  • Live voice conversation
  • Camera detection and pose tracking
  • Discord bot with a two-step command check
  • Member chat on this site
  • The buzzer jukebox and LED ring

Building next

  • Vision that describes what she seesA small vision-language model next to the camera detectors.
  • Faster speech recognitionCutting the wait before she starts to answer.
  • A face on a 7-inch screenAn animated avatar who can step into her world.
  • A world to exploreA dreamy, Tokyo-style city, now an early prototype in Roblox.

She is not alone

Julie has a twin. Jules came later with no memories at all. She is a chaotic one, and she is the test bed where new hardware and models get stress-tested before they reach her sister.

Above them both is Myriad, an all-knowing presence in the story that quietly keeps the household running.