01
Makes things up
Plausible but false answers. Confusing an impressive demo with a tool you can rely on for decision-making.
Loxyn · French deeptech
The organizational sovereign brain.
All your company's knowledge, transformed into a living graph. Each piece of knowledge becomes a node: sourced, dated, permissioned, verified. From document to action — never invented.
The observation
Enterprise AI has a problem of trust.
Generative AI tools impress in demos but disappoint at work. They cannot be trusted to make decisions—and the bill rises with every query.
02
Ignores permissions
No concept of who has the right to see what. Knowledge mixes, confidentialities overlap.
03
Doesn’t know what’s current
No memory of time. An old version contradicts the new one, with no way for the tool to know which one matters.
The building block
Our unit is the fact. Not the document.
Where others split documents into chunks of text, Korela operates at the level of the fact. The brain is the living graph of these nodes—it grows, self-corrects, and remains consistent.
Origin
Where this fact comes from.
Date
Since when it has existed.
Rights
Who can see it.
Truth state
Still true, or contradicted.
Links
What it connects to.
The concept, never the computation.
Sovereign by design
Your data never leaves.
Sovereignty isn’t a checkbox — it’s the architecture. Korela is hosted on your premises (or on a trusted cloud) and can operate with AI models running locally. Your contracts, your emails, your knowledge: nothing is sent to a third party, nothing trains someone else’s model.
You retain control — and compliance.
Your perimeter
Data, AI model, graph — everything stays inside.
100% sovereign
The brain
Four functions. A single architecture.
Above the living graph of atomic nodes, Korela executes four cognitive functions — each derived from the same architecture, none added as an external module.
I
Organize
All of an organization's heterogeneous knowledge, normalized into sourced, unitary facts and interconnected.
II
Search
Natural language querying, with answers grounded in internal sources, traceable, and rights-compliant.
III
Anticipate
Automatic detection of memory risks: undocumented knowledge, fragile dependencies, forgotten deadlines.
IV
Act
Generation of actions grounded in verified facts. AI proposes, humans validate — always.
Functional details and product demonstrations are available at korela.ai.
Research & foundations
Built on science. Not on intuition.
Our knowledge graph doesn’t come out of a hat. Each building block is grounded in peer-reviewed research. Here’s what we rely on—and how it powers us.
Proactive analytics
Summarize knowledge by graph COMMUNITIES → holistic analyses (risks, opportunities, deadlines), not just local search.
From Local to Global: a Graph RAG Approach ↗
Microsoft Research — Edge et al. · 2024
Graph structuring
Partition the graph into GUARANTEED well-connected communities—the very fabric of our reasoning.
From Louvain to Leiden: guaranteeing well-connected communities ↗
Traag, Waltman & van Eck — Nature Sci. Reports · 2019
Entity importance
Identify which entities TRULY MATTER (centrality), beyond mere link counts.
The PageRank Citation Ranking ↗
Brin & Page — Stanford · 1998
Semantic grounding
Link concepts by MEANING, in 100+ languages—executed LOCALLY for data sovereignty.
BGE M3-Embedding: multilingual, multi-granularity ↗
BAAI — Chen et al. · 2024
Entity resolution
Unify variants of the same entity into ONE node (e.g., “Mr. Dupont” = “Marie Dupont”)—no more duplicates.
Extract, Define, Canonicalize (EDC) ↗
Zhang & Soh — EMNLP · 2024
These foundations are public—we cite them. How we combine them, execute them sovereignly, and harden them for real enterprise data? That’s our moat. Details under NDA.
Evaluation
We don’t claim quality. We measure it.
Every evolution of the graph is tested against a benchmark: adversarial documents with a predefined expected graph. We measure entity resolution (B³), cross-document links, and source fidelity in a non-regression framework. A reproducible method, not a promise.
Initial results
On a benchmark of adversarial trap documents, measured across multiple independent runs (extraction is non-deterministic):
Entity resolution — precision 100 % · σ 0
Zero over-merging: two distinct entities are never conflated.
Entity resolution — recall 100 % · σ 0
Multiple spellings of the same entity converge into a single node.
Relations — links between entities 86 %
Room for improvement—acknowledged, measured, not hidden.
B³ metrics (resolution) and F1 (relations), drawn from literature. Zero variance (σ 0) across multiple runs = reproducible, not luck. Adversarial benchmark intentionally limited; large-scale validation (1,000+ documents) ongoing.
Demonstration
A real-world case. The graph in action.
A few documents from a workshop — a lease, an HR file, a supplier, a register. Korela connects them into a graph of facts. Three things then become obvious.
In blue: the point of fragility detected by the brain.
01
It anticipates
The lease is linked to its deadline — the alert is triggered well before 04/30/2027, not on the day itself.
02
It spots the fragility
The dyeing process depends on a single person. The brain flags this as a memory risk.
03
It respects rights
Marie’s file is only visible to authorized individuals — even in a response, no data leak.
Illustrative example — fictional data. No real customer data.
Scientific approach
A demand for research.
We do not confuse a demo with proof. We build on the shoulders of research — and collaborate with it.
Fields involved
01
Semantic memory for robotics
How to represent knowledge that an autonomous agent can use without reinventing? Symbol grounding vs. embeddings, persistent structures, sensorimotor anchoring.
02
Belief revision
Consistent updating of a knowledge base in the face of contradictory or evolving information. Maintaining consistency without losing historicity.
03
Normative reasoning
Native integration of access, compliance, and governance rules within the reasoning layer — not as a post-hoc filter.
Methodological principles
- 01
Falsifiable hypotheses, established before implementation.
- 02
Internal measurements before any external demonstration.
- 03
Systematic comparison with academic state-of-the-art.
- 04
No claim of capability without reproducible proof.
Fields are named; proprietary solutions are not. They are discussed under confidentiality agreements.
The moat
A technological lock. Not a feature.
Anyone can plug in an AI model. Very few can guarantee that an answer is true, up-to-date, and permitted — by design, not by hope.
That's where our core lies: proprietary locks that make trust structural.
They are not on this website. They are discussed — under a confidentiality agreement.
- — Market, competitive landscape, revenue trajectory.
- — Filed or pending patents, intellectual property.
- — Internal architecture, models, measurement methods.
- Sent upon request after NDA.
Detailed under NDA
Trajectory
The same atom, from document to action.
A robot faces the exact same challenge as an organization: to act, it must know what is true, what it is allowed to do, and how to connect a symbol to the real world. This is precisely what Korela can do.
Level 1 · Today
In productionThe organization’s brain.
document → node
Documents, emails, decisions transformed into sourced, dated, rights-managed facts.
Level 2 · Trajectory
ResearchA robot’s memory.
node → action
The layer that enables a robot to act without improvising. Service robotics first, humanoid on the horizon.
This is a research trajectory, not a delivered product. We don’t oversell.
Our research
Humanoid robot.
Today, robots mainly understand what they see. We seek to enable them to understand what the company knows, what they are allowed to know, and what they are allowed to do.
This transition from organizational knowledge to physical action lies at the heart of our research.
The team
Loxyn.
French SAS, based in Lyon (TO-LYON Tower). Integrated technical team. Proprietary stack and locks built in-house.
YC
Yoline Comtet
Presidency
CL
Coralie Lebon
General Management
SG
Sébastien Gracia
Technical Direction
FAQ
Frequently asked questions.
Sovereignty, in concrete terms?
Your data never leaves your perimeter. Korela is hosted on your premises (or on a trusted cloud) and can run with models executed locally. Nothing is sent to a third party.
Where is the data stored?
Wherever you decide. We capture nothing.
How is this different from RAG?
RAG retrieves text snippets and stitches them together, hoping for accuracy. Korela builds a graph of facts that know their origin, date, rights, and truth state. Trust becomes structural.
And the robots, seriously?
This is our research trajectory, not a marketing promise. The knowledge block—true, dated, rights-aware—is exactly what a robot lacks to act without fabricating. Service first.
How can we get in touch?
Contact us via the form at the bottom of the page. We communicate in writing, at your pace.
Contact
Write to us.
Korela is already running on real corporate data — now is the time to get on board. A single message is enough; no mandatory intro call — we’ll exchange in writing first, then move forward together.
Received by the Loxyn team.