DAK by Arzisoft

From a borrowed GPU to our own box

I started learning AI in 2022. Four years later a box on my desk in Beirut runs it without asking anyone's permission. It began as a hobby.

  1. 2022

    I started learning AI.

    No plan and no course. I used it and read everything I could find.

  2. The teachers

    I learned from people who teach in public.

    Each gave me a different piece. The list is further down.

  3. The terminal

    Bash changed how I work.

    Short commands replaced long clicks. Everything got faster.

  4. Two years ago

    Arzisoft began as a hobby between partners.

    I'm into AI solutions, hardware and putting AI into hardware, and I'll tinker with anything open source. My partner is into 3D printing and design, and does our web and design too. Client work paid for the learning.

    See the case studies
  5. The bill

    The better I got, the more the tokens cost.

    I wanted that problem gone, not managed. So I learned to run the models myself.

  6. First local run

    My first local model ran on a partner's graphics card.

    From home, over SSH, on an RTX 4070 across the city. It's my partner's PC. His 3D printing hobby is Beirut Dynamics, our sister workshop, and it's printing our prototype's case now.

  7. Jetson Orin

    Then a machine of my own.

    Small, quiet, a few watts. It showed me how much a little box can do.

  8. AMD, then Mac mini

    One machine became several.

    An AMD mini PC, then a Mac mini with 48 GB. Each took the jobs it was best at.

  9. The turn

    Hardware, not software.

    AI is taking over software. For now, hardware is the future. I put the AI into the machines. My partner designs and prints what holds them.

  10. Proto 1.0

    A pile of leftovers that worked.

    No case. Borrowed and spare machines, wired over SSH on a balcony. It proved this was possible at home.

    Concept render of Proto 1.0.
    Concept render of Proto 1.0.
  11. Proto 1.2

    The setup I use every day.

    Four machines, around 20 open models, our own harness. Its case is going to print.

    CAD render of the Proto 1.2 case, before printing.
    CAD render of the Proto 1.2 case, before printing.
  12. DAK-1

    Then the first product.

    One mainboard in our own case, 128 GB of memory. In progress. DAK is short for Dākhilī, داخلي, Arabic for "local".

    Concept render of DAK-1.
    Concept render of DAK-1.

What I learned along the way.

From people who share what they know for free.

  • Bash

    Living in the terminal.

  • ICM

    Folders and Markdown files in place of a framework.

  • Directing AI

    Guiding it the right way, so it builds what you meant.

  • Building and shipping

    Security, backends, software that reaches people.

  • Local AI

    Open weights, fine-tuning, retrieval.

  • Harnesses

    Everything around a model that makes it useful.

Tell us about your office.

Headcount, the kind of files you handle and what you'd like it to do. We reply within a business day with a first recommendation.

Request a quote →