Apply
DA
APPLY AGENCY / APPEARANCE

See us through your bias

Pick the colour you would rather think in.

↑ ↓ moves · Enter selects · Esc closes.

We stand on shoulders. And at the bar.

Same Apply Agency. Here are the people we owe a nod, and the bits we have to answer for ourselves.

Apply Agency Original
Apply Agency’s own look. Kind. Skilled. Hardworking. With black as the workhorse and a little colour in the personality.
Tokyo Night
A tribute to Omarchy, DHH and the community. The theme was created by Enkia; the palette here is inspired by folke’s TokyoNight port. They took the night shift. We brought the colours to work.
Sea swim & oat latte
Our own blend of sea air, blue ink and soft corners. For anyone who was in the sea at six this morning and would quite like to mention it.
Pistachio with a misprint
Our own palette with warm paper, green tones and Georgia in the headings. Like a small design magazine that has been given access to production.
Lars’ Friday theme
Our own cocktail, mixed for Lars: blue and white with a nod to F.C. København, lime in the glass and FREITAG by Magnus Millang on the stereo. Lars picked the theme. The calendar will have to fall in line.

Under the bonnet: Geist, Geist Mono and Geist Pixel from Vercel, plus Georgia as a system font. Adapted for Apply Agency with Codex. We take responsibility for the names ourselves.

 Contact
AI & agents

Example · Private AI

Private AI on our own models, and frontier models when the task calls for it.

We have bought a cluster of NVIDIA DGX Spark and are setting it up in our own server room on a closed network. Here open models run fully private, and here we can keep training them. Together with EU hosting at Scaleway and frontier models from OpenAI, Anthropic and others, that gives three routes to a model. The portal, the agent and the CMS stay the same, whether you run Umbraco, Sitecore or Optimizely. The task decides the route.

Diagram: CMS, portal and agent point through an AI profile to either the DGX Spark cluster, Scaleway or the frontier providers Your solution Three routes to a model CMS Umbraco · Sitecore · Optimizely Portal Search and answers Agent Data and actions AI profile per task Same code. Switch in the profile. Private DGX Spark cluster Own server room · closed network Open models · own training EU Scaleway, Paris EU hosting · no data stored Open and European models Frontier OpenAI · Anthropic · others Directly from the provider or via Azure · Vercel · Cloudflare
Every task has an AI profile that points to the model the task is allowed to use. The code in the CMS, the portal and the agent is the same wherever the model runs.

The starting point

Some data must not leave the building. Other tasks win on the best model.

There is a big difference between letting a model draft a meta description and letting it read member data, case files, contracts or source code. The first can live anywhere. The second often must not leave the organisation, and certainly not the EU.

That is why we do not want one provider for everything. We want to point each task at the model that fits the data, the price and the quality, without rebuilding the solution. That is what the cluster, the Scaleway agreement and the frontier providers give us together.

Three routes

Private, in the EU or at the frontier providers.

01

Private. Open models in our own server room.

The DGX Spark cluster sits with us on a closed network. The models are open, the weights live on our own disks, and prompts and answers are never sent out of the building. This is where the data that must not leave belongs: member data, case files, contracts and source code. In a trade union the membership itself is sensitive data, so there it covers the whole member system.

How it holds up

  • For member systems, case files, contracts and source code
  • No call leaves the network, and no provider logs anything
  • Fine-tuning on your data, and the model becomes yours
  • A fixed price per month instead of a price per token

02

EU. Open and European models at Scaleway.

When a task needs to scale beyond our own server room but still stay in the EU, we use Scaleway in Paris. Open models and European models such as Mistral run there under European law, and data is not stored by default. We built the Umbraco.AI provider for Scaleway ourselves, so it is a switch in a profile, not a project. Sitecore and Optimizely take the same route through the same API.

How it holds up

  • A European provider under European law
  • For content, documents and enquiries that may leave the building but must stay in the EU
  • No storage of prompts and answers by default
  • Our own Umbraco.AI provider, so Umbraco can switch without code

03

Frontier. OpenAI, Anthropic and others when quality has to be top.

For open and fast tasks where the data is not sensitive and the best model wins, we use the frontier models from OpenAI, Anthropic and others. Directly from the provider, or through Azure, Vercel or Cloudflare when the organisation already has agreements, quotas and security there.

How it holds up

  • OpenAI, Anthropic, Google and others, chosen per task
  • Directly or via Azure, Vercel or Cloudflare
  • Pay per call at the provider's list price
  • For public content, SEO copy and drafts without personal data

The DGX Spark principle

A small cluster in our own server room, wired directly to your CMS and the other services.

NVIDIA DGX Spark is a small machine with the Grace Blackwell GB10 chip and 128 GB of unified memory. One of them runs open models with up to around 200 billion parameters. Link several together and they share the work and can run larger models and more tasks at once.

The cluster serves the models through one OpenAI-compatible API on our closed network. Umbraco.AI, the portal, the agent and the search index talk to it the same way they would talk to OpenAI. The same goes for Sitecore and Optimizely, where we build the connection if the platform does not have one itself. The difference is that no data leaves the building, and the training happens on the same cluster.

Closed network · own server room Closed network · own server room DGX Spark cluster DGX Spark GB10 · 128 GB DGX Spark GB10 · 128 GB DGX Spark GB10 · 128 GB More nodes as needed ConnectX-7 · 200 GbE Own training Fine-tuning on your data Open models Llama · Mistral · Qwen · gpt-oss Local API OpenAI-compatible No calls to the internet Encrypted tunnel CMS Umbraco · Sitecore · Optimizely Portal Hybrid search · answers Agent MCP · actions Other services Documents · code
The DGX Spark cluster serves open models through one local API. The CMS, portals, agents and other services connect through an encrypted tunnel, and training happens on the same cluster.

Across projects

The same models, in several places in the work.

The models are not a project of their own. They sit inside what we already build: in the editor's CMS, in the portal's search, in the agent with access to data and in the operations that follow. Here is where we have written about it.

Under the hood

Built on standards, so the route can be switched.

NVIDIA DGX Spark

GB10 Grace Blackwell Superchip, 128 GB of unified memory and up to a petaflop of AI performance per machine. Several machines link over ConnectX-7 and share a model or a queue of tasks.

OpenAI-compatible API

The models are served in the same API format the frontier providers use. Umbraco.AI, Sitecore, Optimizely, portals and agents switch between the cluster, Scaleway and OpenAI by pointing a profile somewhere else.

Closed network

The cluster can only be reached from our own network and from the solutions we connect through encrypted tunnels. The models have no route out to the internet.

Own training

We fine-tune open models on your documents, language and rules on the cluster. The result is a model you own, which can run with us, with you or in the EU.

What we do not promise

Private is not the same as best. We choose per task.

  • Open models on our own hardware are not always the best. For some tasks the frontier models clearly win, and then we say so.
  • A closed cluster does not make data legal to use. Data processing agreements, purpose and permissions still have to be in place.
  • The cluster grows as needed, but it is built for selected tasks and training, not to replace a cloud for the whole organisation.

Shall we look at where your data is allowed to run?

Tell us which tasks you want AI for and which data is involved. Then we will show you which route they can take, and what it costs.

Write to Kristian
Kristian Corr Thrane
Kristian Corr Thrane

Senior Software Developer and Head of Technology

Kristian is co-founder and technology lead at Apply Agency. He is the one setting up the cluster, and the one to ask about models, networks and training.

kristian@apply.agency