Science and engineering in the same room
Most AI projects fail at a seam. The modelling team does not know why the measurement is noisy; the domain team cannot say whether the model's confidence means anything. Work crosses that boundary by email and loses something every time.
We are built to not have that seam. The same people who train the models have doctorates in the fields the data comes from - materials science, physics, chemistry, biology, statistics - and have published in them. When a prediction looks wrong, the argument about why happens in one room, between people who can each carry both halves of it.
What that means in practice
Domain-level, not vendor-level
We can read your literature, argue with your method section and understand the physics or biology behind your measurement - so we can tell you when the data cannot answer the question, before anyone builds a model on it.
Rigour as a habit
Uncertainty quantified, baselines honest, claims checked. Scientific training is mostly the discipline of not fooling yourself, and it transfers directly to systems people have to trust.
Engineers who ship
Research instincts with production habits: containers, pipelines, monitoring and rollback. A result that only runs on a laptop is not a deliverable.
We are a small, senior team in Belgium, and we work as consortium partners, subcontractors and direct suppliers - with the university network and EU-project experience that comes from having spent years on that side of the table.
Things you can check
Claims about rigour are cheap. These are not:
MicrostructureDB
Our own analysis platform for microstructure imagery - the productised form of the foundation models we deploy in consulting.
A published experiment
One day, one foundation model, an unseen domain - including what did not work. Presented at the IIW workshop at OCAS.
A position on record
Our response to the European Commission AI Office's Expert Forum on Frontier AI, published in full.
What we do
Data & ML
Vision, language and prediction - with calibrated uncertainty and explanations attached.
Science
Materials and life sciences modelling, and the products that came out of it.
Cloud
Deployment and the software around it, so a model stays useful after the pilot ends.
AI Trust & Security
Red-teaming and audits for LLM applications and agentic systems.