Sovereign AI and federated learning
Your process data never leaves the site. Your sites still learn from each other.
A pilot that dies in the security review has wasted everyone's year
Cloud-first industrial AI runs into three walls at once: data residency law, customer contracts that forbid process data leaving the plant, and the simple fact that a cell recipe or a weld parameter set is a trade secret. IT security is right to block it, and a pilot that dies in a security review has wasted everyone's year.
Our default architecture assumes none of that is negotiable. Inference runs on an edge device at the machine. Training and monitoring run on hardware you control. Where several plants would benefit from learning together, federated learning sends model updates to an on-premise aggregator and never the recordings.
The deliverable that decides whether this gets approved is not the model. It is the architecture and data-flow documentation your security team reads, mapped to IEC 62443 where OT components are involved, with every external connection named and justified.
What is included
Architecture for on-premise or on-device training and inference with no cloud dependency; data governance and access design; federated learning across sites or partner companies (secure aggregation, model versioning, monitoring); open-weight models and open tooling where possible so you are not locked to a vendor; documentation for IT security and compliance reviews, mapped to IEC 62443 where OT components are involved.
An architecture that survives a security review
Data flows, trust boundaries, update paths and failure modes documented before anything is installed.
On-premise training and inference
No cloud dependency anywhere in the loop, including model updates and monitoring.
Federated learning across sites
Secure aggregation, model versioning and per-site drift monitoring, so a shared model improves without any site exposing its data.
Open weights and open tooling
Wherever they meet the requirement, so the system outlives the relationship with us.
Four stages, each with an output you can check
Governance mapping
Who owns which data, what may cross which boundary, and what your contracts with your own customers already commit you to.
Architecture
Edge, on-premise and, where relevant, federated topology, with the documentation written for the reviewers, not for us.
Pilot at one site
Full loop proven on a single plant: collect, train, deploy, monitor.
Federation
Additional sites joined, aggregation and versioning monitored, drift tracked per site rather than in aggregate.
Sectors that buy this most often
What clients ask before they start
Is federated learning worth the complexity for two plants?
Usually not. It earns its keep from roughly four sites upward, or where sites belong to different legal entities and cannot pool data at all. For two plants we would normally say so and keep it simple.
Do you need remote access to our network?
No. We can work entirely on site, or on an air-gapped copy of the data that your team exports. Remote access is a convenience, never a requirement.
What about the models themselves — are they yours or ours?
The trained weights are yours. We do not train across clients, and your data never contributes to another company's model.
Ask about Sovereign AI & federated learning
Tell us the asset or process and what goes wrong. We answer every enquiry within 48 hours on working days, and the first call is free.
- NDAs signed before the first call if you prefer
- Anonymised or synthetic samples are fine to start
- Your data is never used to train models for anyone else
Services clients combine with this one
Ready to find out if it works on your machines?
Start with a free 30 to 45 minute discovery call. Bring the problem, the recordings you already have and your questions. We will tell you honestly whether sensing-based AI is the right tool and what the next step would cost in days, not months.