Paris · London · USA · China · Africa · on-premise
Build theautonomouscompany.
Software that waits for a ticket is the old stack. Cybernecs designs AI agents, Cerebros, and sovereign infrastructure so operations run, report, and stay under a named human when the stake is high.
HITL on financial, destructive, or security actions
Headquarters · Paris
The shift
Software was the first phase. Autonomy is the operating layer.
Enterprise tools sit idle until someone clicks. AI-native companies route work across a living graph — sales, finance, engineering, support — and execute with a mandate, not a chatbot in the corner.
01 — Systems
What we actually build
Not a model wrapped in a landing page. Six layers that turn an existing organisation into an AI-native operation — with traces, approvals, and a place to run.
01
Cognitive agents
Task-specific, state-aware clusters that plan multi-step work, call internal APIs, and stop when the mandate says so.
02
AI-native software
Business systems designed around reasoning, retrieval, and live data — not another CRUD form on a legacy ERP.
03
Agentic surfaces
Interfaces that read intent in plain language and assemble the next screen, instead of burying work in static menus.
04
Cerebros
The nervous system: ERPs, messaging, code, and finance telemetry connected to one reasoning kernel.
05
Private deployments
On-premise and air-gapped inference when intellectual property cannot sit on a vendor cloud.
06
SecOps with HITL
Continuous inspection of model calls, tokens, and data leaving the perimeter — halt before leak, not after.
Cerebros · one graph for the whole firm
02 — Cerebros
Your company as a living knowledge graph
Sales, treasury, engineering, support, and logistics stop living in separate tabs. Cerebros maps them into one operational memory so agents reason on the same facts.
Sales and pipeline vectors — qualification with a written mandate
Treasury and cashflow — forecasts, never silent payments
Software operations — PRs and services with a named approver
Client intelligence — triage with an exportable trace
Sovereignty — cryptographic policy on every model call
Financial, destructive, or security actions always submit a proposal. A human named in the mandate signs, or it does not run. Cerebros · On-premise.
03 — Agents
A workforce that reports
Agents are not a demo chat. Each one has an identity, a memory horizon, and an approval gate. The interesting part is what they are allowed to do without asking — and what they must never do alone.
Human-in-the-loop · named humans keep the last word
04 — Method
Six steps. No theatre.
We observe the real operation first. Then we model, connect, and deploy agents with bounded autonomy. Automation comes after the mandate is signed — not before.
01 Observe
Capture the work as it actually happens
APIs, inboxes, ledgers, and the exceptions people handle by phone.
OutputTelemetry map
02 Model
Company graph and twin
A sovereign knowledge graph you can simulate before you touch production.
OutputExecutable model
03 Connect
Harden the pipes
Tokens, retrieval, and policy gates. Nothing leaves without a trace.
OutputGoverned connectors
04 Deploy
Agent workforce
Purpose-built agents with a written mandate and a named human on the gate.
OutputLive runtime
05 Automate
Move the repetitive work
Routing, analysis, and drafts run continuously. Exceptions stay human.
OutputBounded autonomy
06 Evolve
Learn from operations
Agents refine from telemetry. Policy and HITL stay in the loop.
OutputOperating rhythm
Guides
Where to go next
Six crawlable pages, one intent each. Same facts as the identity cards — no stuffing, no invented ROI.
COS-01, REV-09, DEV-44, AUD-88 — plus machine JSON at /.well-known/agent.json.
05 — Work
Where the architecture lands
Hospitality, freight, and industrial operations — places where a wrong price, a missed slot, or a leaked document actually costs money. We design the system; we do not invent client percentages on this page.
Hospitality & travel
Direct reservation and concierge layer
Replace fragmented booking desks with a private agent that reads inventory, rates, and house rules — and never invents a price. Human confirmation on upgrades and exceptions.
Documents, berths, and inland transfer stop living in three inboxes. The site calculates from your rules; the agent proposes; a named operator releases the truck.
Stack · sovereign inference · edge twin · HITL release
Presence
Paris headquarters. Built for operators worldwide.
Research in France. Delivery in the United Kingdom, the United States, China, and Africa. On-premise when data cannot leave the country.
USAChinaAfrica
Paris
Research, model architecture, Cerebros
London
Delivery, English-speaking operators
USA
Delivery and private deployments
China
Sovereign / in-country runtime
Africa
Industrial and logistics operations
On-premise
Your racks, your weights, your mandate
06 — Contact
Let’s build your next operating system.
Talk with a systems architect. We review the stack you already have and name where agents create measurable work — not a slide about “AI transformation”.
