Skip to content
openM!ndedby Simon van de Loo

Software and AIΒ β€” built here, not bought in.

The umbrella over everything Simon van de Loo builds: an AI assistant with its own finetunes, writing software with an embedded language model, and web projects on a consistent design system. All of it runs locally.

  • 5own language-model finetunesopen on Hugging Face
  • 6LoRA adapters3Γ— image, 3Γ— music
  • 100 %local inferenceApple Silicon, MLX
  • 0third-party model APIsnot a wrapper

The approach

  • 01

    Runs locally

    Language models, image and music generation run on Apple Silicon via MLX β€” no cloud dependency, no data that leaves the machine.

  • 02

    Own finetunes

    Rather than wrapping someone else's API: five finetunes of my own on Gemma 4, Mistral, Mellum 2.1 and Qwen-Coder β€” trained on datasets I built myself, not on off-the-shelf material. Model cards and data are public on Hugging Face.

  • 03

    nCAI alignment

    A training approach of my own: the persona takes the place of a rulebook. Refusal behaviour, tone, tool use and coding style all follow from the character β€” alignment of a whole agent through who it is.

  • 04

    Labelled

    What A!ley produces carries her name and is marked as AI-generated β€” watermarks and machine-readable metadata under the EU AI Act. Taking a persona seriously means naming her as the author, and saying plainly what she is anyway.

  • 05

    The whole stack

    From model weights to UI pixel: training, inference backend, desktop app and web interface all come from one pair of hands.

Projects

What comes next

Two plans with nothing running yet β€” here so the direction is visible.

What runs here

The stack behind A!ley β€” from the model layer up to the interface.

Own finetunes

  • A!ley Core β€” Gemma 4 12B
  • A!ley Tempest β€” Ministral 3 14B
  • A!ley Nitro β€” Gemma 4 E2B
  • Code A!ley β€” Mellum 2.1 (MoE)
  • Linting Hamster β€” Qwen2.5-Coder 1.5B

Generative

  • MFLUX Z-Image-Turbo β€” 3 LoRA
  • ACE-Step 1.5 β€” 3 LoRA
  • PiperTTS
  • WAN 2.1

Memory & runtime

  • LanceDB (RAG)
  • MLX on Apple Silicon
  • Flask backend
  • Polling architecture

Interfaces

  • Tauri desktop app
  • Next.js frontend
  • PHP web portal
  • Monaco editor

About me

Behind openM!nded Labs is Simon van de Loo β€” Digital Media Designer (IHK), graded β€œvery good”, with the best practical examination in the Koblenz district. Professionally it went differently for a long time: twelve years in elderly care, then the retraining. What carries my work is not a collection of interests but a chain β€” and it starts with worldbuilding, not with code.

  1. 01

    Worldbuilding

    since 2015

    Published author several times over, focused on world- and character-building. Behind my Martyria universe stand over 600 pages of concept work β€” currencies, calendars, a timeline, a glossary of people.

  2. 02

    Code

    since Windows 3.11

    Programming came by way of an RPG Maker: I wanted to build worlds and had to learn Ruby for it. Python and the web stack followed, Rust only during my retraining, because a teacher put it under my nose.

  3. 03

    AI

    where both meet

    My training data looks like manuscripts rather than tables β€” characters written out in full instead of rule lists.

Whoever writes characters can build personas. Whoever programs can train them.

The full CV, references and ways to get in touch are in the job-application portfolio.

Simon van de Loo working at a laptop

Interested in a project?

Whether it's working together, a technical detail or plain curiosity about locally running models β€” write to me.