03 / 09 — Services

Custom AI Models & Training

In short

A custom AI model makes sense when an off-the-shelf model does not know your terminology or is not accurate enough for your tasks. We develop models for companies and train them on the industry you need, for example law, medicine, engineering, retail or logistics. In most cases we start from an open language model and fine-tune it with your data; for classification, forecasting or image recognition we also build smaller models from scratch. The model belongs to you and runs on your own hardware or in a cloud of your choice.

How it works.

The typical data flow — from input to result. Every station is replaceable and adapted to your systems.

01Your data
02Preparation
03Training / fine-tuning
04Evaluation
05Your own model

Glowing blue fibre-optic cables symbolising data transfer

What we deliver

Custom AI models for your company, trained on data from your industry.

01Assessment whether a custom model is needed or RAG with an existing model is enough
02Preparation of training data from documents, tickets, databases and domain texts
03Fine-tuning of open language models (e.g. with LoRA) for your industry and terminology
04Custom models for classification, forecasting and image recognition
05Evaluation with your real tasks, measured before and after
06Deployment on premises or in the cloud, with an API and regular retraining

A look at the code.

This is what an excerpt from a typical project looks like — readable, commented and without magic. That is exactly how we hand code over to your team.

PyTorchHugging FaceLoRAPythonTensorFlow.jsOpenCV

train.pyExample

from qc.train import Dataset, FineTune, evaluate

daten  = Dataset.aus_ordner("kanzlei/schriftsaetze", format="chat")   # eigene Beispiele
basis  = "open-instruct-8b"                                          # offenes Basismodell
lauf   = FineTune(basis, methode="lora", epochen=3, lernrate=2e-4)
modell = lauf.trainiere(daten.train)

print(evaluate(basis,  daten.test))    # vorher
print(evaluate(modell, daten.test))    # nachher, gleiche Testfragen
modell.exportiere("gguf", quant="q4_k_m")   # läuft danach lokal

What you get out of it.

Knows your terminology

The model learns the terms and style of your industry and your own texts.

Measured, not promised

We test before and after training with the same real tasks.

You own the model

You get the model files and training code, with no per-request licence.

Runs where you want

On your own server, a workstation or a cloud with EU data centres.

Typical scenarios

PyTorchHugging FaceLoRAPythonTensorFlow.jsOpenCVCUDAGGUFllama.cppDocker

03.1

Law & consulting

A language model that knows legal terminology and the firm’s own templates and drafts documents in the firm’s style.

03.2

Healthcare & practices

A model that understands medical terms and structures findings or doctor’s letters, running entirely on the practice’s own server.

03.3

Industry & engineering

A model for maintenance reports and error messages that classifies faults and suggests measures from past cases.

03.4

Retail & service

Classification of enquiries, returns or product data with a small custom model that runs fast and at low cost.

Matching solutions.

Abstract neural network symbolising artificial intelligence

Company knowledge assistant

A chat assistant that uses RAG to search all approved sources. Permissions are checked per document and every answer shows its source.

Answers with evidence
Permissions preserved
Runs locally or in the EU

Private in-house AI platform

A GPU server with local language models and an internal chat interface, with user management and logging and an API for your own tools.

Company data stays in-house
Predictable cost
One interface for everyone

Frequently asked questions

When do you need a custom AI model?

When an existing model does not know your terminology, writes in the wrong style or makes too many mistakes on a fixed task. If you only need knowledge from your documents, RAG with an existing model is often enough and cheaper; we check this with you first.

How much data is needed for training?

For fine-tuning, a few hundred to a few thousand good examples are often enough, for example questions with correct answers or texts in the required style. A model built from scratch needs considerably more; we can say how much once we have looked at your data.

Who owns the trained model?

You do. You receive the model files, the training code and the documentation and can run the model on your own hardware without ongoing licence fees.

Does our training data stay confidential?

Yes. If you want, the whole training runs on your own server or on a server in Germany, so the data does not go to a US cloud provider.

More services

01 / 09

Back to overview

Ready when you are

Let’s talk.

Describe your project in a few sentences — you’ll get an honest assessment of feasibility, effort and the way forward.

Discuss a project info@quell-code.com

E-Mail: info@quell-code.com
Phone: Call-back on request
Response time: usually 2 working days
Languages: Deutsch · English · Türkçe