AI Development

Our principle: artificial intelligence (AI) as controlled processing, not as an isolated answer generator. Where a matter can be calculated or decided by a rule, we calculate and decide. Where language, images, or fuzzy relationships have to be understood, we use AI models.

Machine learning as the core of AI development

Machine learning is the part of artificial intelligence in which a method does not obtain its capability from fixed rules, but learns from examples. It is the core of our AI development: image models, language models, and agents are based on it. We use deep learning, that is, learning with deep neural networks, where images, language, or sensor data have to be understood.

A machine learning project follows a fixed process with us:

1. Collect and prepare data. We review your data, clean and standardise it, and record which version was used for which model.
2. Annotate. Examples are given the information the model is to learn. Automatic pre-annotation speeds up the work; people correct and approve.
3. Select and train the model. We adapt a pretrained model to your task or train a dedicated one, depending on the data situation and the requirements.
4. Evaluate. The model is tested on data it has not seen during training and measured against agreed criteria.
5. Deploy and update. The model runs in the target system, on the server or on an embedded computer. Corrections from operation flow into the next training run.

Where the task requires it, we combine machine learning with classical data analysis, rules, and physical models. We describe applications under Vision AI, GeoAI, Sensor Fusion & Physical AI, and Person Detection.

Structure of an AI pipeline

We break complex tasks down into specialised subtasks and combine the results in a pipeline:

1. Sources. Documents, project data, logs, images, sensor data: text, image, audio, and structured data are processed together.
2. Knowledge layer. Indexing, vector search, and context building provide what is relevant to the question at hand.
3. Analysis. Classical data analysis, rules, statistics, and classification.
4. AI models. Language models, image models, models for audio and text, created through machine learning.
5. Agents. Specialised agents with their own skills and tools, in workflows with verification.
6. Use. Decision, documentation, traceability, and reuse.

Reliability through multiple paths

Where the result matters, several independent AI paths work on the same question. If they agree, reliability increases; if they differ, the result is reviewed. Corrections flow back into the pipeline.

People remain responsible

The AI structures, compares, proposes, and creates transparency. People assess, prioritise, and take responsibility. At critical points, human approval is built firmly into the process.