Signature module · JT Digital
Private LLM fleet
The models run on machines we control, so your material never becomes someone else's business.
Chat · Extract · Classify
Most software that says AI means a round trip to somebody else's API, with your documents as the payload and a meter running for as long as the feature exists. Ours makes no such trip. The open models sit on one machine in one European datacentre. The files they read rest in that same datacentre. The machine has no address the internet can reach. It dials out to the queue, and nothing dials in. Same features, without handing over the material or the meter.
What it does
The useful half of AI, without the handover.
Six things clients actually ask for, all of them running on the same machine.
Answers from your own material
Questions in plain language, answered against your own documents and data, with nothing leaving the machine that holds them.
Documents turned into fields
Invoices, contracts and forms come back as structured data your software can use. No retyping.
Sorted by your rules
Tickets, applications and messages routed the way your team already sorts them, at whatever volume shows up.
Text out of pictures
A vision model reads what is in an image or a video frame: scanned pages, screenshots, signage, stills pulled from footage.
Thousands of files in one run
Long jobs go to a queue and come back as a file of results, so nobody sits watching a progress bar.
Every result traceable
Each batch result carries a record of which model and which weights produced it, so an answer can still be explained months later.
Why it matters
Four things you get back.
Your material never leaves
The model runs on a machine we control. Its interface is bound to the machine's own loopback address, so nothing on the internet can reach it. The machine dials out to our queue. Nothing dials in. There is no third-party API key in the path because there is no third party in the path.
Every application gets its own queue credentials, scoped to its own queues and nothing else.
The models are open-weight, under Apache-2.0 or MIT licences, so nobody can change the terms under you.
We do not train or fine-tune anything on what you send.
Your files never leave either
Every document, scan and image you send is stored in a bucket we control, in the same datacentre as the machine that reads it. There is no file-sharing service in the middle, no vendor upload endpoint, and no copy of your material sitting in somebody else's account. The attachment and the model are neighbours.
The queue carries a reference to the file. What your document contains never rides through the message broker.
Storage, queue and model sit in one datacentre, so a file crosses between them in a fraction of a millisecond. Nothing goes over the public internet.
Results come back to the same storage. They never pass through a vendor's dashboard.
The bill does not move with usage
Capacity is a fixed monthly rental. Ten thousand documents this month and none the next cost the same, so the number in your budget is the number you pay. What moves under load is how long the queue takes. The invoice stays flat, and no vendor can reprice the model in the middle of your project.
No per-token metering, no usage tier to graduate into, no surprise at the end of a busy quarter.
Capacity is finite and shared, so we size it with you before we promise a turnaround.
European, and specific about where
Inference, the queue and the storage that holds your files all sit in one European datacentre. The applications that call them run on a Slovenian host, and the company you would be contracting with is Slovenian. When your lawyers ask where the processing happens and where the documents rest, the answer is one address.
In practice
What is actually running.
Wired into every new project
Every project we start ships with the client for the fleet already in place. Using a model is a function call, no procurement and no new vendor contract.
Proven end to end in production
The batch path runs sealed jobs on the queue: caption, extract and classify, validated end to end in production, each result carrying a record of the model and the weights behind it.
One machine, one model at a time
Capacity today is a single card that also serves video work, holding one model in memory at a time. That is plenty for documents, classification and overnight batches. If your project needs more, we rent more before we promise it, and you will know the number before you sign.
The rest of the head start
Two more we own outright.
Next step
Bring us the data you cannot send away.
Fifteen minutes is enough to tell you whether a private model fits your case, and what running it would cost per month.