Éter
A versioned cryptographic library consumed by Moeris, Nova and future developments: one improvement propagates to every product at once, with no duplicated sensitive code.
Architecture
Sixteen platforms are not sixteen loose projects because they share cryptography, a model factory, automation, data interconnection and a written AI governance constitution.
Components
A versioned cryptographic library consumed by Moeris, Nova and future developments: one improvement propagates to every product at once, with no duplicated sensitive code.
Turns data into versioned, signed datasets, trains models and ships them as a portable package with full traceability of which data produced which model. It is what lets us say the AI is ours.
Detects a real, unattended flaw: keys that look correct but are guessable because the entropy source was broken. Born from a documented theft of 60 million euros in 41 minutes.
A proprietary visual workflow engine with nodes specific to our domains — planning, real estate, legal — that generic platforms do not understand.
The central layer connecting the ecosystem’s applications so the same data is never loaded twice or kept as two different truths.
A written operating constitution for our AI systems: signed identity per module, mandatory logging of every action, human authority able to freeze or revert, and a ban on any system widening its own permissions.
Proprietary engines
Every artificial intelligence capability in the ecosystem is packaged as an independent, versioned engine, reusable by any platform. These are not third-party integrations wrapped in an interface: they are our own code under proprietary licence, which is why an improvement in one of them reaches every vertical that consumes it at once. We group them by family; the engine-by-engine detail, with its declared maturity, is delivered during due diligence.
Inference and in-house models
A local inference gateway with several adapters, a unified training pipeline and specialisation adapters that fine-tune an expert model without sending anything to the cloud.
Consumed by: The whole portfolio
Verified knowledge
Retrieval over proprietary sources with a faithfulness verifier, plus a post-generation check that contrasts what the model wrote against authorised sources before letting it out.
Consumed by: Cybersecurity · Urban planning · Real estate · Expert reports
Decision and learning
Deep reinforcement learning agents, tree search to reason down several paths before committing to one, and adversarial red-versus-blue training.
Consumed by: Energy · Cybersecurity
Monitoring and anomaly
Anomaly detection combining several methods, security event correlation with open rules, and observability of how models behave in production.
Consumed by: Cybersecurity · Energy
Domain computation
Numerical engines written for one specific problem and free of commercial dependencies, starting with the power-flow solver with bus switching and dynamic topology.
Consumed by: Energy
AI governance and safety
Ethical decision patterns with automatic veto and an auditable trail, and code generation with security validation and isolated execution. This is the part that makes what MCP-NOVA writes as policy enforceable in software.
Consumed by: The whole portfolio
Sovereign runtime
An in-house replacement for the usual container orchestration, built only on the operating-system kernel and the standard library, to deploy where third-party software is not accepted.
Consumed by: Infrastructure
The maturity catalogue deliberately excludes non-productive engines — cloned templates never implemented, and code with no consumers — which sit in quarantine by documented decision and do not count as part of the inventory. We prefer a smaller, certain count to a larger, arguable one.
A vertical product consumes Éter’s cryptographic core, the models packaged by Crucible and i7n’s workflows, publishes its data through INFINITY Brain and operates within the limits written in MCP-NOVA. An improvement in any of those pieces reaches every vertical without rewriting product.