janus 4 products with code and tests 🇪🇺 Built in Romania

One engine · eleven domains

Data you can actually use — and proof that it exposes no one.

Name: JANUS — sovereign synthetic data
Brand: CAI Technology

JANUS produces synthetic data for model training where the real data is too scarce, too sensitive or impossible to share. Every delivery carries three independent proofs, one of which matters most: that nobody in the original set can be recognised in what was generated.

synapse-eu.eu ↗ Discuss a pilot →

0.82

reconstruction fidelity, measured across 2,600 volumes

4 of 11

products with code and tests; the rest are specifications

3 proofs

utility, fidelity and privacy — or the batch is rejected

Your side

generation runs on the client's own infrastructure, in the EU

The problem

The data models need is exactly the data nobody can hand over.

Three constraints that combine and stall entire projects.

There is not enough real data

A diagnostic algorithm needs thousands of annotated cases; a hospital produces a few hundred a year. A pathology with forty cases annually will never yield enough real data, however long you wait for it.

Real data cannot be shared

The exact information a model needs is the protected kind: patients, a grid's demand patterns, industrial process parameters. The vendor cannot build without it and the holder cannot hand it over.

Anonymisation alone is not evidence

"We removed the names" is not a proof. Without a measurement of re-identification risk, an ethics committee has nothing to rest on — and rightly refuses to.

The common recipe

The same three beats, whatever the domain.

Whether the data is a tomography, an electricity demand curve or a radio signal. What changes between domains is only the kind of data and what we condition on.

Compress 01

Keep the structure, drop the bulk

A model learns to hold real data in a far more compact form while preserving what matters. For a scan that means the anatomy; for an electrical grid, the demand patterns and how they depend on each other. This is what makes generation possible at scale.

Generate 02

Ask for exactly what you lack

A second model creates new compact representations, refining noise step by step, guided by what you ask for: a lesion on this artery, a grid failure at peak hour, a vibration anomaly on the front bearing. Each result arrives with its label, because both are produced together rather than separately.

Verify 03

Three proofs, or nothing ships

That it resembles reality statistically. That a model trained only on the synthetic data works on real data. And that nobody in the original set can be recognised or reconstructed from what was generated. Fail one, and the batch is rejected.

In development

Four products with code and tests.

One vertical taken deep, and three transversal products that apply to any data.

Advanced development

JANUS Med

Train medical AI without touching a patient.

Generates CT and MRI volumes that look and behave like real ones, together with the segmentations models need, without corresponding to any existing patient.

You can ask for exactly what you need: a lesion of a given size, on a given artery, ruptured or not, in a patient of a given age. That is what solves the rare-case problem. Every delivered set comes with a validation report: how close it is to reality, whether a model trained on it works on real patients, and proof that no original case can be reconstructed.

For: Medical software vendors · hospitals · clinical research · regulators

In development

JANUS Core

The trust layer.

A synthetic dataset without evidence is just a large file. Core turns each delivery into something a compliance officer can accept on argument rather than on faith: it keeps the record of what was generated, from which source, with which parameters and under which guarantees, and seals each set with verifiable proof of provenance.

It administers who has access to what, keeps the complete operations log, and issues the certificate that travels with the data when it moves on. The difference between "we generated some data" and "I can demonstrate at any audit how it was generated, and that it exposes no one".

For: Security directors · data protection officers · compliance teams

In development

JANUS Data

Protects data in motion.

Most leaks do not come from spectacular attacks but from ordinary flows: an export to a supplier, a file sent by email, an integration between two systems. Data sees where sensitive information travels and masks or removes it at the moment it moves — on bulk upload, through a connector or API, or live from a database.

Detection is automatic, by type of information, not by rules written by hand for each field. Sensitive originals stay put and what travels onward is already clean. It is also what feeds the generation engine: whatever enters training has already passed through anonymisation.

For: Security and data teams · organisations with compliance obligations

In development

JANUS Edge

The enforcement point.

Data protection policies are worth something only if someone actually applies them, at the point the data passes through. Edge is that point: the gate information travels through between systems, applying the rules set in Core before anything moves on. Without it, a policy stays a document.

It has been through an adversarial security review — not merely tests confirming that it works, but deliberate attempts to break it, with everything found fixed before any real use. It is designed to run inside the organisation, with no external dependencies.

For: Infrastructure and security teams that must guarantee nothing sensitive leaves unintentionally

What is measured

The numbers we have, and the claims we do not make.

0.82 reconstruction fidelity, across 2,600 volumes.

The reference open model (NVIDIA MAISI) reaches 0.84 — carrying four times the parameters and trained on roughly four times the data. The generator is in its first full training run. Medical imaging is the first domain taken all the way, so it is the only one with measured figures.

Built but not calibrated. The quality gate exists and works — it correctly rejected a deliberately weak model — but its thresholds are provisional and have to be calibrated on real data before they mean anything in public.

What we do not claim. That the generated data has been clinically validated, that a medical device trained on it has passed any certification, or that the proposed verticals have been demonstrated. Seven of the eleven products are specifications with a shared engine proven on the first domain — not things you can install.

The real differentiator is not image quality. Models that generate medical data exist. Very few can demonstrate numerically that from what they generated you cannot get back to a real person. For regulated data, that is the argument that opens an ethics committee's door.

Specification

Seven more, on paper for now.

Product specifications on the same engine. No code, no delivery date, and nothing here is for sale. They are listed because the engine is domain-agnostic, and these are the domains where the shortage of shareable data hurts most.

Specification

JANUS Geo

Would generate Earth-observation imagery and cadastral records, including the hard-to-capture scenarios — floods, sudden vegetation change, unauthorised construction — with the classifications already attached.

geospatial · cadastre · agriculture

Specification

JANUS Energy

Would generate demand, generation and grid-behaviour series, including the stress scenarios an operator cannot produce on request and would never share.

grids · generation · demand

Specification

JANUS IoT

Would generate realistic telemetry, normal and abnormal, with the anomalies already labelled, so predictive-maintenance models see every kind of fault without waiting for it.

sensors · buildings · telematics

Specification

JANUS OT

Would generate industrial process and control data, including fault conditions, so algorithms can be built and hardened without the plant exposing how it runs.

industry · critical infrastructure

Specification

JANUS UAS

Would generate the sensor signatures used for drone detection and tracking, including combinations reality supplies only rarely.

defence · site security · airports

Specification

JANUS RF

Would create signal and spectrum data covering exactly the cases real recordings miss — the ones an operator meets once every few years.

defence · telecom · spectrum monitoring

Specification

JANUS Text

Would produce documents, reports and records that read like the authentic ones, with the domain's terminology, without any real person's information.

regulated documents · health · legal

Pilot

What you can start today.

The transversal part — Data and Edge — is what can be installed now, on your real traffic. The rest of the family grows behind it.

A two-week discovery sprint — integration architecture plus a demonstration on your real traffic, in both modes: anonymise and block.
Policies per data category, written jointly with your DPO and CISO — not copied from another client.
Installation on your infrastructure, on-premise or EU private cloud, with a reproducible build.
An admin and DPO console with the full decision log and per-category statistics.
A DPO runbook — how to answer a transparency request, an incident, a NIS2 audit.
Support through the first 90 days to calibrate detection against your own corpus.

Usual rhythm: two weeks of discovery, four to six weeks of productive pilot, then a decision made together.

Next step

Bring a representative set, and we measure on your data.

The honest way to find out whether this works for your domain is not a brochure. Tell us what you are trying to train and what you are not allowed to share.

Open synapse-eu.eu ↗ Write to us →

Frequently asked questions

How does JANUS relate to synapse-eu.eu? +

SYNAPSE-EU is the platform; JANUS is the family of products that delivers it. The same relationship as between scanope.com and ARTEMIS: a public product site, and the product running behind it. The full family description, with each product's status, lives at synapse-eu.eu/janus/.

What can I actually use today? +

Four products have code and tests: Med, the medical imaging vertical, which is the most advanced; Core, the trust layer; Data, which anonymises information in motion; and Edge, the enforcement point. The other seven — Geo, Energy, IoT, OT, UAS, RF and Text — are product specifications. We do not present them as deliverables, because they are not.

What figures have you measured? +

Reconstruction fidelity reaches 0.82 across a corpus of 2,600 volumes, against 0.84 for the reference open model (NVIDIA MAISI) — which carries four times the parameters and trained on roughly four times the data. The generator is in its first full training run. The quality gate exists and works, but its thresholds are provisional and must be calibrated on real data.

Has the generated data been clinically validated? +

No, and we do not claim it has. Nor do we claim that a medical device trained on it has passed any certification, or that the proposed verticals have been demonstrated. We are an engine proven on the first domain, not a certified medical product — and the difference matters far too much to blur.

What, then, is the real differentiator? +

Not image quality. Models that generate medical data exist; very few can demonstrate numerically that from what they generated you cannot get back to a real person. For regulated data, that is the argument that opens an ethics committee's door — not a fidelity score with two decimals.

Does our data leave the perimeter to be processed? +

No. Generation and anonymisation run on your infrastructure, on-premise or EU private cloud. The component that protects your data does not, itself, take your data out of the perimeter — neither to an external model nor to CAI Technology.

How do I start a pilot? +

Write to us through the contact form with the subject "JANUS pilot". We reply within 24 working hours. The honest way to find out whether this works for your domain is not a brochure: bring a representative set and we measure on your data.