AI models on your data – running on your infrastructure.
We train and run AI for your processes: open-source LLMs fine-tuned on your data, RAG over your documents or classic ML models – privately hosted, with no data sent to US APIs.
$ bytefactory ai --data ./documents ✔ Documents indexed ✔ Base model selected: open-source LLM $ bytefactory train --method lora epoch 1/3 ··· ok epoch 2/3 ··· ok epoch 3/3 ··· ok evaluation ··· better than base model $ bytefactory deploy --region eu hosting ··· your infrastructure
Your data stays your data.
AI that uses your knowledge without it ever leaving the building.
Fine-tuning
Open LLMs such as Llama, Mistral or Qwen trained on your language and tasks with LoRA/QLoRA.
RAG & documents
Answers from your own documents – with sources instead of hallucinations.
Private hosting
On your hardware or in the EU – GPU servers or Kubernetes, no US API.
What's included.
- Data check & use-case assessment
- Fine-tuning of open LLMs (LoRA/QLoRA)
- RAG over your documents
- Classic ML & vision models
- Evaluation with your test cases
- Operation on your own or EU infrastructure
How a project runs.
- Step 1
Use case & data
We check goal, data and value – honestly, before you invest.
- Step 2
Prototype
A quick prototype shows what works with your data.
- Step 3
Training & evaluation
Fine-tuning or a RAG pipeline, measured against your test cases.
- Ongoing
Operations
Privately hosted, monitored and retrained when needed.
Custom LLM fine-tuning or RAG?
Anyone considering custom LLM fine-tuning should first clarify what the model has to do, because that decides whether fine-tuning, RAG or both make sense. With RAG (retrieval-augmented generation), your documents are indexed. For each question, the system looks up the matching passages and passes them to the language model as context. Answers can be backed by sources, and new documents are available without retraining.
With fine-tuning, the weights of an existing model are adjusted using your own examples. That pays off when the model must reliably master a specific style, format or domain language. LoRA and QLoRA train only small additional weights, which considerably reduces the hardware needed compared with full fine-tuning. To make facts reliably available, on the other hand, fine-tuning is less suitable than RAG. Often the combination works best. For classic tasks such as classification or image recognition, custom ML models are an option as well.
The order matters: first we check goal, data and value in the data check, honestly, before you invest. Then a quick prototype shows what works with your data. Only after that come training or the RAG pipeline, measured against your test cases. Without test cases of your own, there is no way to tell whether a model is actually better than the base model. For RAG, your existing documents are enough. For fine-tuning, a few hundred to a few thousand good examples are often sufficient. Which open models such as Llama, Mistral or Qwen come into question depends on language, task and hardware. The license terms of the models also differ and should be reviewed before use. Before training, data should be cleaned and, if it contains personal data, pseudonymized as far as the purpose allows.
One possible field of application is document processing. How far it can be automated depends on the document types, the quality of the templates and the error tolerance in the process. In general, document-processing workflows consist of capturing incoming documents, recognizing the document type, extracting relevant fields and passing them on to the next system. Where mistakes have consequences, a review by people belongs in the process.
Data protection and private hosting
Anyone using an external AI interface transmits inputs to the provider, meaning questions and, where applicable, excerpts from documents. If these contain personal data, questions arise about data processing and transfers to third countries. That is exactly why we run models on your own hardware or on infrastructure in the EU, without calls to US APIs. Your data stays your data.
Technically, the models run on GPU servers or on Kubernetes. How much hardware is needed depends on model size and load: smaller open models can run on a single GPU, larger ones need several. A Kubernetes environment has the same requirements for security and operations as any other application; the page on Kubernetes consulting describes what that involves. If your data and systems still sit in a previous environment, a cloud migration into an EU region may make sense first.
With RAG, there is another point: permissions. Users should only receive documents through the system that they are allowed to see otherwise. Depending on the use case, the EU AI Act can also be relevant. Request logs can also contain confidential content and belong in the operations plan. After the start, the model is privately hosted, monitored and retrained when needed.
Example: document automation
For a non-profit association in Cologne, a tailored AI workflow was built in 2025 that automates document processing and makes everyday work faster for the team.
Do you have a similar use case? Describe it briefly: in the free 30-minute intro call, we clarify with the engineer who will also run your model whether and how it can be implemented.
AI document automation
A tailored AI workflow that automates document processing and makes everyday work faster for the team.
Frequently asked questions.
Do we need fine-tuning or is RAG enough?
If the model should reproduce knowledge from your documents, RAG is usually enough. Fine-tuning pays off when it must reliably master a specific style, format or domain language. Often the combination works best.
Does our data leave the company?
Not if you don't want it to. We run models on your own hardware or on infrastructure in the EU – without calls to US APIs.
Which models do you use?
Open models such as Llama, Mistral or Qwen, depending on language, task and available hardware. For classic tasks like classification or image recognition, custom ML models too.
How much data do we need?
For RAG, your existing documents are enough. For fine-tuning, a few hundred to a few thousand good examples are often sufficient. The data check tells you what is realistic.
What does a custom AI model cost?
It depends on the use case. We start with a fixed-price analysis phase – data check and prototype – before larger investments. If you then commission the project, we credit it.