Phase 2: Build. The assessments are deliberately short. They take weeks, have a fixed price and end in a document you own. Most of our work comes after an assessment, or instead of one when you already know what you need, so an assessment is not a prerequisite.
Three kinds of work
Custom models and systems
Vision, language and prediction systems, built the way how we build describes: a foundation model adapted to your domain, a fast model on top, calibrated uncertainty on everything, and a deployment that learns from its own mistakes. The scope runs from a first prototype to a system your team operates.
Scientific and industrial R&D
Work in materials, microscopy, life sciences and process data, where the model has to respect the physics or biology and the result has to survive a reviewer. We often work alongside your own scientists, and sometimes in place of a hire you cannot make.
Research partnerships
Horizon Europe, VLAIO, ESF and similar. We join consortia as the AI partner, act as subcontractor, or help write the proposal. We have sat on that side of the table ourselves, through AID4GREENEST and AI4MI.
How a project starts
- 1
Scoping
Usually an assessment, sometimes just a conversation. Either way the first deliverable is a written scope with a price and what "done" means.
- 2
Prototype
The smallest thing that answers the real question, which typically takes a few weeks. It ends with a measured result, even if that result is "this will not work".
- 3
Build
The work is split into milestones, and each one ships something usable. You see progress in working software rather than in status reports.
- 4
Handover
Documentation, tests and the monitoring that tells you when the model is drifting. It is built for your team to run without us, and we stay available if you would rather we ran it.
How it is priced. When the scope is clear, as it usually is after an assessment, a project is fixed price against milestones, each of which ships something you can use. When the scope is not yet clear, we work to a time budget with a ceiling and check in with you well before we reach it. That is better than inventing a number we would both regret. Research work follows whatever the funder's rules are.
Phase 3: Run
Building the system is not the end of the work. A model in production drifts, the data behind it changes, and the people who understand it move on. The last conversation of every project is therefore about who runs it afterwards. That can mean a handover with documentation and tests, monitoring that tells you when something has shifted and retests at regular intervals, or it can mean we keep operating it for you. Either way, the decision is made before the build ends.
What we are careful about
We will tell you early if it is not a fit. A project we should not take costs you more than it costs us, and a scope we cannot deliver damages the one thing we sell. If the right answer is "hire someone in-house" or "this is a software purchase, not a model", we say so.
Confidential by default. Most of this work is under NDA and stays there. We publish our methods but not client details, and we ask before naming anyone.
Where to go next
Start with an assessment
Not sure yet? A short review is the cheapest way to find out what a project should be.
Science
The domains this work comes from, and the products it produced.
Cloud and deployment
What running a model in production involves.