Evaluation dataset
Together with your domain experts we assemble examples with a known correct outcome. This dataset is the benchmark for every model variant, and it remains your property.
Adapting AI modelsAdapting AI models
Adapting AI models
where it actually pays off.
We adapt language models to your tasks, through fine-tuning, small dedicated models, or prompting and structured outputs. Measurement decides which route fits.
Book a free AI opportunity review
45 minutes, no strings attached. You leave with a written shortlist.
What it is about
Adapting AI models
Many fine-tuning requests can be solved without fine-tuning. Good prompting, retrieval over your documents, and strictly defined output formats often bring an existing model up to the required level. That is faster, cheaper, and easier to maintain.
Fine-tuning earns its place in clear cases: when a smaller model needs to handle a narrow task as well as a large one, when a fixed format or domain vocabulary has to be reliable, or when high volumes make cost per transaction the deciding factor.
Every decision rests on an evaluation set built from your real cases. We use it to compare the options and see whether an adapted model genuinely beats the simpler route.
What we build
The building blocks
Adapting API models to your use case
Prompting, in-context examples, retrieval, and structured outputs against a fixed schema. This turns a general-purpose model into a tool for your task, with no training involved.
Fine-tuning open-source models
Where the numbers justify it, we fine-tune an open-source model on your data, usually with efficient methods such as LoRA. The result runs on your own or EU infrastructure.
Small specialised models
Classification and extraction often need only a small model that is fast and cheap to run. Typical tasks are routing enquiries, recognising document types, or reading fixed fields.
Comparison and recommendation
You receive a side-by-side comparison of the variants by quality, latency, and cost per transaction. It ends in a reasoned recommendation, which may well be not to fine-tune.
Approach
From the first conversation to operations
- 01
Free AI opportunity review
In 45 minutes we clarify what the model needs to do and whether adapting it is even a candidate. No obligation and no cost.
- 02
Measure the baseline
We build the evaluation set and measure how well existing models perform without training. Any adaptation has to beat this baseline.
- 03
Fixed-price adaptation
Prompting, fine-tuning, or a small dedicated model: the chosen approach is delivered as a package with a fixed price and acceptance criteria measured on the dataset.
- 04
Operations and retraining
When your data changes or better base models appear, we test again. We only retrain when the measurements show an improvement.
By industry
Where this applies in practice
The industry overview lists concrete use cases. Detailed pages are available in German.
- Machinery and plant engineeringUse cases: Technical documentation
- Logistics and freight forwardingUse cases: Reading and checking freight documents
- Medical practices and outpatient centresUse cases: Documentation drafts for physician approval
- Real estate and property managementUse cases: Property listing texts
- Retail and e-commerceUse cases: Product copy and PIM data maintenance
- Industry and manufacturing
- Utilities and municipal energy providersUse cases: Pre-sorting market communication exceptions
- Insurance brokersUse cases: Drafting advice documentation under IDD
- Construction and architectureUse cases: Bill of quantities and GAEB analysis for estimating
- Staffing agenciesUse cases: Drafting job advertisements
- Hotels and hospitalityUse cases: Analysing reviews and drafting responses
- Agriculture and foodUse cases: Checking labels and allergen declarations against recipes
Related services
What often comes with it
- On-premise LLMsWe set up open-source language models so that your data stays inside your organisation or the EU. With permissions, audit logs, and controlled updates.
- RAG knowledge assistantYour staff ask questions in plain language and get answers drawn from your own documents. Every answer points to the passage it came from.
- Document AIWe automate your incoming mail: documents are recognised, read, checked, and passed to your systems. Uncertain cases go to a person for approval.
FAQ
Questions about Adapting AI models
When is LLM fine-tuning worth it?
When a narrow task comes up very often, when a fixed format or domain vocabulary must be reliable, or when a smaller model should replace a large one for cost or data protection reasons. For questions about your own documents, retrieval is usually the better route.
Fine-tuning or RAG: which is better?
They solve different problems. RAG gives the model access to current knowledge from your documents, with sources. Fine-tuning changes the model behaviour, style, or output format. For knowledge questions RAG is almost always the first choice, with fine-tuning added if needed.
How much data do you need to fine-tune a model?
Depending on the task, a few hundred to a few thousand good examples are often enough to fine-tune an existing model. Quality matters more than volume. How much your case needs becomes clear from measurements on the evaluation set.
Can you build a custom AI model for us?
Yes, in the sense of an adapted open-source model or a small specialised model for classification or extraction. Training a large language model from scratch is not something we recommend to mid-sized companies.
Do we own the fine-tuned model?
Yes. The trained weights, the training data, and the evaluation set belong to you. The base model licence still applies; we check and document it during selection.
What does it cost to adapt a language model?
It depends on whether prompting is enough or training is needed, and on the effort required for example data. With methods such as LoRA, compute for training is usually the smaller part. After the baseline measurement we quote a fixed price per package.
Adapting AI models: let's find out in 45 minutes whether it pays off for you.
Book a free AI opportunity review
45 minutes, no strings attached. You leave with a written shortlist.