How we put AI to work
We work on all kinds of AI projects, from simple chatbots to complete AI platforms.
Our AI projects draw on machine learning, computer vision and generative AI. We work independently of vendors with proprietary and open models, including open-weight models and models we have trained ourselves. The largest model is not always the right choice: smaller models running locally can often offer important advantages. The right choice depends on the use case, appropriate use of resources and our clients' requirements for data sovereignty.
Chatbot with RAG
A chatbot should do more than give plausible-sounding answers: it should find the right information. We used Retrieval-Augmented Generation (RAG) for this. The system retrieves relevant content and uses it as the basis for its answers.
Text-to-image fine-tuning
The image model needed to produce results that better matched the desired subjects and qualities. We trained an existing model on a curated dataset, making the generated images more relevant and convincing.
Solar panel detector using satellite images
Solar panels can be hard to distinguish from similar surfaces, such as skylights, in satellite images. We developed a system trained on annotated images that reliably detects solar panels.
Extracting data from unstructured text
We built a Kubeflow pipeline that uses different LLMs to process large volumes of text and extract structured information. We wanted not only to get results, but also to know how reliable they were. A validation dataset let us measure the error rate and reduce it through targeted adjustments.
Developer training
Anyone building with AI should understand what is behind it. Our training covers the history and fundamentals of machine learning and AI, the development of natural language processing (NLP), embeddings and how LLMs work. We also cover today’s LLM tooling, such as vector databases and MLOps, for building applications and running them in everyday use.