A consortium across 20-plus countries wanted to train a breakthrough model for early detection of genetic disease. The blocker was not technical, it was the law: privacy rules like GDPR and HIPAA forbid moving patient data out of the country it was collected in, so a single global database was illegal and small local models were too weak to matter.
A model this powerful needs the widest possible range of patients, but the obvious approach, pool everyone's scans and genomes into one global database, was flatly illegal. Data-sovereignty rules meant patient data could not cross the border it was collected in, and trying to build a central data lake was blocked before it began.
Even setting the law aside, the logistics were impossible. Moving enormous medical imaging files, MRI and CT scans, between continents ran straight into bandwidth limits that made a central collection unworkable.
And the partners were not set up for it. There was no uniform, managed infrastructure inside the hospitals and labs to run complex algorithms locally, so the compute could not simply be pushed out to where the data lived either.
Train inside each country, and move only what the model learned, never a patient record.
We stood up a network of distributed server nodes placed locally in each country, sitting inside the boundary of its own regulator, so the processing happens where the data legally has to stay.
A federated-learning architecture trains the shared model separately inside each country, on that country's local patient data, so no one ever has to gather the data into one place to learn from it.
A private data fabric moves just the model's mathematical weights between countries, never real clinical data, so the global model updates and improves while every patient record stays home.
Because no sensitive record ever crosses a border, GDPR, HIPAA and every local data-sovereignty rule hold by construction, not by a policy bolted on afterwards.
The heavy imaging never moves; only the algorithm's learnings, a few megabytes, travel between sites, which turns an impossible cross-continental transfer into a routine one.
The unified global model still learns from the full genetic diversity of more than 20 countries, capturing the breadth that makes rare-disease detection work, without ever centralising the data.
Compliant by design, and more accurate for it.
The consortium met every data-sovereignty and privacy law, cleared the international project with the regulators, and started within months. The unified model reached unprecedented accuracy on rare diseases precisely because it learned from 20 countries' genetic diversity.
Because the heavy scans stayed put and only the model's insights travelled, international data traffic fell over 95%, and not one piece of sensitive patient information ever left its home country.
Pooling sensitive data to train a model is the assumption that breaks on privacy law and on bandwidth. The value here is inverting it: keep every record where the law requires, train locally, and share only the mathematics the model learned. Compliance becomes a property of the architecture rather than a promise. It is the same privacy-first discipline behind our medical-anonymisation work, aimed at training across borders rather than de-identifying a single dataset.
A medtech firm's clinical data was locked away by privacy law. An AI de-identification engine masked the direct identifiers while keeping the clinical findings, with human review guaranteeing zero leaks and meeting HIPAA and GDPR in full.
Read the case → Data readiness · InsuranceAn insurer's AI hallucinated because its data was fragmented. A rebuilt cleaning and metadata pipeline with permission-scoped retrieval and an 85% confidence floor made it dependable for field staff.
Read the case →
Reef TRH