AI for companies that don't want to hand over their data.

Synblob is a small studio for applied AI. We run workshops, build prototypes, and turn the ones that work into tools your team uses every day. Where we can, we use open models that run on your own servers or with a European provider.

what we do

Most of what people ask us for falls into one of three categories.

Workshops. One or two days in which your team works with current AI tools themselves. Afterwards everyone knows what the tools are good at and where they get things wrong. You also leave with a list of ideas that could be worth pursuing in your company.

Prototypes. We take one of those ideas and, within a few weeks, build a version you can actually try. For example a search across your internal documents, an assistant for customer service, or image generation trained on your own products.

Implementation. If the prototype holds up, we develop it together with your team into a tool that runs in daily use. It comes with documentation, so you don't depend on us afterwards.

how we work

It starts with a conversation in which we try to understand where things are stuck. Sometimes AI isn't the right answer, and if that's the case we'll say so.

For projects we agree on scope, price and timeline up front. You see the current state every week and can try it out yourself. We work in German and English, on site or remote.

independence

Many AI services run on servers in the US, and your data goes there. For some uses that's fine. For customer data, contracts or engineering drawings it often isn't. Open models such as Mistral, Qwen or Flux are now good enough to run yourself, on your own hardware or with a hosting provider in Germany. We've been working with these models since 2023 and have a good sense of what they can do and what they cost to run.

If it doesn't matter where the data is processed, we use whichever tool fits best.

examples

A few projects we worked on before Synblob.

Collections sorted by similarity. An open-source tool that lets museums and archives show thousands of images at once, arranged by time, theme and visual similarity. An image model computes the similarity. Institutions in Germany, Switzerland, Sweden and the US use it for their collections. The same approach works for product catalogues, image archives or technical drawings.

Image generation on dedicated hardware. A handheld camera that sends photos to a GPU server, where an image pipeline re-creates the scene. It was shown at festivals and conferences in Germany and Belgium. The same kind of pipeline can produce product images in a consistent style, trained on your own photos.

Data that stays with its owner. For a publicly funded research project we built visualisations for a desktop app. People used it to explore their own YouTube and TikTok data on their own computer before deciding whether to donate it to research. Tools that handle sensitive company data can work the same way.

background

We've been building software and data visualisations for more than fifteen years, for newsrooms, museums and research institutes among others. From 2023 to 2025 we worked on a research project at a German design school about making AI understandable and tangible. It led to installations shown at festivals and conferences, and to courses for design students. We have also led the rollout of AI tools at a software company. So we know both sides: building the tool, and what has to happen afterwards so that people actually use it.