Plants, branches, subsidiaries, countries: each with its own system, owner and jurisdiction. Copying it all to one store is a transfer to justify, a store to protect and a year to wait.
Each plant, branch or subsidiary holds its data under its own system, its own owner and often its own jurisdiction. Getting read access to all of them takes a programme of its own.
Copying personal or confidential data to a central store creates a new transfer to justify, a new store to protect, and a new residency question in every country involved.
Terabytes of logs or images from twenty sites over ordinary links. By the time the lake is full, the question has changed.
A training client at each site learns from local data. A coordinator combines the learning into one model and sends it back. No copy, no transfer, no lake. The sites keep their data and their rules, and the group gets one model that knows all of them.
Three parts: the network of clients, the data at each site, and the privacy of what travels.
Clients at the sites, a coordinator in the middle.
Aligned without being moved.
What travels reveals nothing.
Four phases, from the sites on a list to a model at every one of them.
Sites, data, rules, the question.
Clients, coordinator, connections.
Rounds across the sites.
The model everywhere, new sites joining.
One model across the group, without the migration and without the transfer.
No data lake to build first. Clients installed, schema agreed, first round run: weeks, not the year a central platform takes.
Data stays in the country and the system it lives in. The compliance question per site is answered by the design.
Patterns from every plant, branch or subsidiary, measured at each site against a model trained on its own data alone.
| Side by side | Central training | Decentralized training |
|---|---|---|
| Data movement | Everything to one store | None; updates only |
| Privacy | A second copy to protect | Records stay on site, updates masked |
| Compliance | A transfer per jurisdiction | Satisfied by the design |
| Working across organisations | Needs a sharing agreement | Needs a protocol |
| Who controls the data | The central platform | Each site, still |
Usually close, and it is the comparison that matters only when a central store is possible. When it is not, the comparison is with what each site can train alone, and the decentralized model is better by a margin we measure per site on local held-out data and report before release.
Groups with data in several plants, branches, subsidiaries or countries: manufacturing, retail chains, banking and insurance groups, health networks, public bodies with regional offices. And consortia of separate organisations that want one model without sharing data, which is federated learning across companies.
Every connection is authenticated and encrypted. Updates are combined with secure aggregation, so the coordinator sees only the sum. Where the guarantee requires it, differential privacy noise is added. Each update is bounded against poisoning. Each site logs its rounds for its own audit, and the protocol is reviewed by each site's security team before the first round.
A machine inside its perimeter able to run the training client on its local data: for most tabular and log data an ordinary server, for images or large models a GPU. Plus an outbound connection to the coordinator. The coordinator runs wherever the group allows, on premise or in a private cloud. We size it in assessment from the data and the model.
We are a small team of senior specialists. We pick the right model and the right layer, and we build the least machinery that does the job. You get a call with an engineer, not a sales deck.
