FROM REAL PROBLEMS

Working software.

I build local AI and workflow software, working with AI coding assistants from requirements and system design through implementation, debugging and release checks.

ATLAS COMPANION0.18.7 / WINDOWS
Windows 0.18.7. Scripted test data, local core disconnected. Captured 10 October 2026.
Windows 0.18.7. Scripted test data, local core disconnected. Captured 10 October 2026.

LOCAL AI. WORKFLOW SOFTWARE. BELGIUM.

I see the problem. I build the fix.

SCROLL TO EXPLORE Open to junior roles & freelance SELECTED WORK / 2026
Explore Atlas
FIRSTHAND EXPERIENCE.
PERSONAL INITIATIVE.
BUILT IN BELGIUM

01 / FROM THE TERMINAL FLOORINDEPENDENT PROTOTYPE

MULTILINGUAL DRIVER ASSISTANCE

A better way
to ask for help.

Prototype / working demo

International drivers need help with everyday problems at vehicle terminals. My first-hand experience of that work inspired this independent prototype.

A flat battery. No fuel. Damage. A car that is not where it should be. Or a map they cannot read. A phone number on their paperwork is only a starting point. Explaining the problem is difficult when the driver and desk worker do not share a language.

THE PROTOTYPE, ON SCREEN

A conversation.
A vehicle.
A clear ticket.

I built a WhatsApp assistance prototype that guides drivers in their own language. They choose what they need, photograph or type the vehicle identification number, confirm it, and select the issue. The request arrives on a dispatch board for the appropriate team.

1 min 51 sec · Recorded prototype · Dutch interface

Recorded 10 September 2026. Company-neutral presentation of a local simulator, with demonstration data and mock yard locations.
Technical decisions behind the conversation

Keep the request consistent across interruptions.

The prototype keeps conversation state while the driver moves between help, map and vehicle steps. An issue selected early is retained until the VIN is confirmed. Confirmation buttons carry the VIN they refer to, so a previous button cannot confirm a different vehicle.

The simulator and dispatch board use the same local ticket store. Regression tests use isolated stores and mocked external requests. This demonstrates the workflow locally; it does not establish readiness for a live terminal deployment.

Read the English walkthrough
  1. Ask for help. The recording starts the guided simulator. The driver receives a menu with emergency, vehicle problem, desk and terminal-map options. This recording uses Dutch.
  2. Identify the vehicle. The driver chooses a vehicle problem and sends a VIN-plate photo. The system reads the vehicle identification number and asks the driver to confirm it.
  3. Describe the issue. The driver selects battery / will not start. A ticket appears on the dispatch board, with the issue, vehicle identifier and demonstration location.
  4. Give dispatch the context. The board shows the technical queue and desk/scanner context. Opening the ticket reveals the VIN, vehicle information, photo and controls to take or resolve the request.
01 / THE DRIVER

Help in their own language.

A guided conversation with clear choices, a VIN photo and confirmation. Drivers can describe their problem without having to explain it over the phone in English.

02 / THE DESK

Context in one place.

The vehicle, issue and supporting photo arrive together. The board provides technical, desk/scanner and emergency queues so staff can see what needs attention.

03 / THE NEXT STEP

A pilot on a real terminal.

This is a personal prototype with a working local demonstration. I can run a small pilot with the people who do the work. A live terminal connection is not built yet. That would be part of the pilot.

Next step If you run a terminal, I can pilot this with your team on a real site. The recording is a local simulator. Connecting it there would be the work of the pilot.

In development

ANOTHER TERMINAL WORKFLOW

From paperwork
to reviewed data.

My freight-document parser takes a photo or upload of a CMR consignment note and uses a vision model to extract vehicle and shipment details. Desk staff can review the fields, correct them and confirm a structured record.

The capture-station flow checks document placement and image quality before taking a photo. Uncertain fields go back to a person for review. It is still in development; a live terminal-system connection has not been demonstrated.

  • Node.js
  • Vision AI
  • Document capture
  • Field validation
  • Human review
Explore the freight parser ↗

02 / LOCAL INTELLIGENCERUNTIME · TOOLS · INTERFACES

Meet Atlas.

A local AI companion.
A shared runtime behind it.

Atlas connects a Windows companion and an Android client to a local Python runtime. Recent work focuses on the boundaries between a model's proposal, owner approval and a checked result, with saved phone tools alongside chat. Voice and wider reliability remain active work.

Watch the Atlas recordingATLAS COMPANION / 2026
  • Workflows across local apps. A coordinator carries tasks through planning and result checks. Local app adapters connect tools, and a recorded Image Studio workflow generates an image while handing GPU memory between models. The case study explains what was tested and what remains open.
  • Native inference, on my hardware. Recent Atlas development includes an experimental executor for Qwen and original Gemma 4 12B weights, using PyTorch and custom CUDA kernels. The managed llama.cpp path remains available. Model execution and memory use are part of the design.
  • Always-on voice. Wake word, hands-free answers out loud, streaming TTS with a natural voice ladder. Echo and barge-in are the hard, half-solved parts, and I can tell you exactly why they are hard.
  • From desktop to phone. The early phone app was a remote microphone. The native Android client now brings chat, voice and approvals to the desktop brain. Its release and remaining device checks are covered in the build story.
  • Checked outcomes and owner approval. For declared file changes, the lab work joins scoped approval, a saved checkpoint and independent content checks. A passing selected test alone is not enough.
  • Small tools beyond the chat. A prebuilt Pocket Tool was kept on Android and retained its saved state after an app restart with the desktop service stopped. One checked workflow, with broader tool creation still being developed.
  • A face, not a textbox. An animated face and HUD, plus a droid avatar with 64 sprite poses. It looks alive because a companion should.
  1. Act I

    Jarvis

    I wanted my own Jarvis: a companion that remembered, helped me build software, worked on my PC and shared everyday tasks.

  2. Act II

    The OS detour

    Autonomous work kept meeting a sensible boundary around protected system files. I tried building an operating layer with AI at its centre, then realised the scale of that project.

  3. Act III

    Atlas Companion

    I returned to the companion I wanted at the start. It lives beside my normal Windows apps and carries the local brain, tools, memory and safety work from the OS experiment.

Ten documented lab scenarios cover specific outcomes and failure cases. Broader reliability and everyday use remain active work.

03 / ATLAS ENGINENATIVE INFERENCE · MEMORY · CUDA

THE RUNTIME IS ITS OWN PROJECT

Inside the model.
Down to the kernel.

Companion is the interface. Atlas Engine is the experimental inference runtime: loading model weights, managing memory, executing CUDA kernels and generating tokens. It is where much of the deeper systems work happens.

Explore Atlas Engine
ATLAS ENGINE / MEMORY PATHARCHITECTURE DIAGRAM
01 / MODEL FILEQwen3.5 · 35B A3BGGUF · Q4_K_M · published model weights
Validated loading
02 / SYSTEM MEMORYCompressed expert banksExpert weights stay in RAM
Bounded transfers + reuse
03 / GPU EXECUTIONCUDA + expert cacheModel state, attention and token generation
Simplified 35B execution path on the test PC: 64 GB system RAM and a 12 GB GPU. The complete model is not resident in GPU memory.
MODEL PARAMETERS
35BQwen3.5 MoE · about 3B active
GPU MEMORY
12 GBBounded GPU residency
SYSTEM MEMORY
64 GBThe recorded test PC

Recorded native comparison, 26 October 2026: 17.4 → 26.4 tokens/s with the same measured output. Loading and prompt processing are excluded. Read the method and limits

04 / THE APPROACHHOW I BUILD

From requirements
to verified behaviour.

I work with AI coding assistants on architecture, implementation and debugging. AI writes much of the code. My contribution is defining the behaviour, working through trade-offs, testing what was built and deciding what needs another pass.

01

Define the system.

I start with the user workflow and its failure cases. With AI support, I work through which components are needed, what data they exchange and where permissions belong. Those decisions become requirements for the implementation.

02

Implement, then investigate.

I break features into smaller changes, try them in the application and bring concrete failures back into the next iteration. That includes UI behaviour, model configuration, hardware limits and the connections between components.

03

Check the result.

I use demos, regression tests and release checks to challenge a change. For Atlas, that includes checking actual file contents and whether existing files were protected. I distinguish a passing test from a feature proven in everyday use.

Define the behaviour.
Verify the result.

I am self-taught and still building my engineering knowledge. I am looking for a junior role where I can contribute this practical experience and learn from an experienced team.

DESIGN DECISIONSTHREE CONCRETE EXAMPLES

What sits behind
the interface.

The interesting work is in the boundaries: what the model may do, what the software checks and when a person needs to decide.

01 / ATLAS

A result needs evidence.

The model proposes an action. The runtime checks its scope and approval. Declared file outcomes need independent content and test checks before they can be marked verified.

Explore the outcome checks

02 / TERMINAL ASSISTANCE

A conversation with state.

A driver can ask for a map, return to a request and confirm a vehicle. The flow must remember the right issue and reject an old VIN confirmation.

Read the workflow decision

03 / FREIGHT PARSER

Extract, validate, review.

A vision model proposes document fields. Schema and VIN checks catch invalid output; a person reviews the record before it is confirmed.

See the validation approach

05 / SELECTED WORKBUILT. USED. REFINED.

From the workbench.

Tools I use, prototypes I can demonstrate, and experiments that taught me something. Open a project to see the interface, my contribution and where it stands.

06 / BEYOND THE TERMINALSAME CURIOSITY. DIFFERENT FREQUENCY.

Off the clock.

MUSIC & VIDEO

Also known as Misterhaze.

The tracks are AI-made under my direction: my stories in the lyrics, my call on the genre, generated in Suno. The videos are the part I craft myself, cut in DaVinci Resolve with my own toolkit doing the tedious parts.

BASED IN BELGIUM

Dutch native. English fluent.

Based in Belgium. Haven is aimed at Dutch-speaking firms; everything else I build and learn in English.

07 / YOUR MOVE

Let’s build
something useful.

I build and test software and automation around concrete problems I come across. Especially repetitive tasks, cumbersome processes and things that make you think: surely this could be simpler?

Does your business run into a problem like that? I’d like to hear how you handle it today. I’m happy to explore where software might help.

I develop these projects myself, with help from AI tools. I can show what already works, how I’ve tested it and where there’s still work to do.

wardgeys.dev@gmail.com
Connect on LinkedIn

Write in Gmail ↗ Open your email app ↗

Belgium · replies in Dutch or English

Atlas Companion

Existing Atlas recording. Historical footage, not proof of the current release.