01
Framing before model
The problem stated, the expected benefit, and how you will know it worked. Measured against the simpler option, because the simpler option often wins and that is a result worth having.
Artificial intelligence
AI you can host yourself, not AI you rent by the token
Open-weight models on infrastructure you control, with the licence you are accepting named out loud and the way back out documented before the first line of code.
We treat AI as something that augments a team, never something that replaces one. The need therefore comes before the model: we write down the problem, weigh it against the duller options — better documentation, ordinary software, a changed process — and only then ask which model earns its place. When one does, it runs where your data already lives.
Where we work
02
Model selection and evaluation
Open weights by default, and a European publisher wherever one is competitive: Mistral covers both language and speech under Apache 2.0. Each candidate evaluated on your data rather than on a leaderboard, and its licence read first — the strongest image weights, FLUX, are non-commercial, and Mistral's own code model stopped being open at all. Open weights is not open source, and an open model is not always one you are allowed to ship.
03
Sovereign deployment
On your own servers, or at a European provider such as OVHcloud or Scaleway. GPU sizing from measured load, inference that stays inside your perimeter, and a documented path to the next model when this one is retired.
04
Local and on-device inference
Quantised models running on the workstation, for the cases where nothing should leave the machine at all. Confidentiality by construction rather than by contract, inside a resource budget we actually measure.
Where it pays off
01
Assisted engineering
Code suggestions, refactoring, generated documentation and test scaffolding, inside the editor your team already uses. What comes out is a draft, and a named person signs it.
02
Review support
Static analysis widened to what a linter cannot see: an abstraction nobody uses any more, error handling that changes shape halfway through, a migration applied to four files out of five. It flags; it never approves.
03
Operations and logs
Anomaly detection across logs and metrics, and configuration checked against what the system actually does. We do not promise incident prediction. We promise fewer alerts nobody reads.
04
Business workflows
Ticket classification, technical-writing assistance, and analysis of user feedback at a volume nobody reads by hand. Each one measured against the manual baseline, and dropped if it does not beat it.
From audit to autonomy
The same engagement shape as the rest of our work: at the end, you are the one operating it.
- Audit of the need and the constraints — data, regulation, existing systems, and the simpler option we have to rule out first.
- Selection of models and technologies, with licensing and hosting examined before capability.
- Infrastructure built and documented, from GPU sizing to the exit conditions.
- Training for your teams on what the system does, what it cannot do, and how it fails.
- Skills transfer: your people operate it while we watch, not the other way round.
- Support afterwards, flexible, for as long as that is useful and no longer.
And what we refuse to do
Our responsible-AI charter is published, versioned and printable. It names what we will not do: client data in a public service without explicit approval, a consequential decision handed to a model, a generated output presented as verified fact. Read it before you brief us.
How an engagement runs
A framing phase short enough that abandoning the idea stays cheap, a written assessment that belongs to you, and a deployment your team can operate without us. If AI is not the answer, the report says so.
Talk to us