One governed intelligence cluster.
Eight product layers, seven control planes, and a governed execution path every production action must pass through — on infrastructure you control.
The intelligence cluster architecture for governed execution.
NeuroCluster is organized as eight product layers within one governed intelligence cluster — operated through seven control planes and enforced by a governed execution path every action must pass through.
Experience Layer
Console, workspace, and admin surfaces for operators and end users.
Unified UX across Intel workspace, platform console, and tenant admin — designed for governed AI operations at scale.
App Layer
Application catalog, deployments, and business workflow orchestration.
Deploy vertical and horizontal apps with promotion pipelines, environment isolation, and tenant-scoped configuration.
Intel Layer
Agents, models, RAG, tools, evaluation, and reasoning systems.
LangGraph cognitive loops, Temporal workflows, Supernova reasoning integration, and governed tool execution.
Data Layer
Lakehouse, knowledge bases, vectors, and semantic context.
Row-level access policies, classification ceilings, embeddings pipelines, and lineage-aware retrieval.
Integration Layer
ERP, CRM, SaaS connectors, webhooks, and external APIs.
Pipedream and Composio connectors, governed outbound actions, and auditable integration profiles.
Compute Layer
Kubernetes, GPU clusters, Firecracker sandboxes, and GitOps.
Distributed inference, model routing, horizontal agent scaling, and sovereign deployment topologies.
Security Layer
Identity, SSO, RBAC, tenant isolation, and secrets management.
Authentik SSO, Vault secrets, network policies, and per-tenant cryptographic boundaries.
Governance Layer
Policies, HITL approvals, audit trails, and evidence packs.
OPA policy gates, EU AI Act readiness, exportable evidence, and procurement-grade audit artifacts.
Every runtime action resolves identity and tenant context, checks permissions and policy, executes in scoped runtime, and emits trace, audit, and evidence events. This is the kernel contract of the cluster.
Agents that cannot take high-risk action without proof.
Build and deploy agents with prompt versions, tool permissions, evaluation gates, and approval workflows — every production binding frozen at promote time.
Governed agents follow a single lifecycle: draft, test, evaluate, approve, stage, deploy, monitor, improve, retire. Production promotion requires evaluation results and policy attachment — not ad-hoc toggles.
The platform proof demo walks a governed enterprise agent from governed data through a high-risk action, policy gate, human approval, and exportable evidence pack.
One gateway for models, tools, and inference policy.
Route open-weight and commercial models through a single control plane — with fallbacks, quotas, payload logging, and MCP endpoint governance.
Enterprise buyers need routing, cost visibility, and policy at the gateway — not just API keys. LiteLLM integration provides unified inference with tenant-scoped quotas and observability hooks.
MCP servers register as first-class connections: owner, ACL, tool list, and PolicyBundle attachment. Coding agents and IDE integrations inherit the same governance model as production agents.
Approved data becomes governed AI context.
All retrieval flows through the Data API — with row-level policies, classification ceilings, knowledge base ACLs, and lineage back to source.
Agents must not call Qdrant or object storage directly. The Data API enforces tenant scope, row_filter access policies, and classification ceilings at retrieval time.
Knowledge bases bind documents, chunk policy, and retrieval ACLs. Data products expose curated gold tables with owners and classification for governed agent workflows.
Your infrastructure. Scoped execution.
Run agents, apps, and sandboxes on Kubernetes with Firecracker isolation, GPU nodes for open-weight models, and secret scopes limited to each workload.
Agent-generated code and sensitive integrations need runtime isolation beyond standard containers. Firecracker microVMs provide ephemeral sandboxes that destroy on completion.
RuntimeSandbox resources bind network policy, secret scope, and deployment stage. High-risk actions execute only inside approved runtime boundaries.
- Kubernetes with Helm + ArgoCD GitOps
- Firecracker microVM sandboxes
- GPU nodes for open-weight models
- SecretScope-bound workloads
Proof for security, legal, and procurement — not promises.
Every production action emits trace, audit, and lineage events. Evidence packs bundle deployment snapshots, approval history, policies, and inference logs for vendor review.
Regulated buyers do not buy guardrails — they buy reviewable proof. EvidencePack is an exportable derived resource: who approved what, with which prompt, model route, and policy bundle, at which timestamp.
Platform audit, Phoenix traces, and OpenLineage events chain together so answers and actions trace back to source data, model, prompt, policy, user, and approval.
Choose the boundary that fits the workflow.
NeuroCluster deploys on your terms — from managed validation environments to sovereign air-gapped clusters. Installation profiles match team size, sector constraints, and procurement requirements.
Air-gapped
- No external network paths
- Offline signed updates
- Sovereign & government workloads
On-premises
- Runs in your datacenter
- Identity-integrated
- Full operational control
European cloud
- EU data residency
- GDPR-first architecture
- Managed sovereign services
Dedicated tenant
- Isolated operating boundary
- HA, Vault, SCIM
- Evidence packs & backup/DR
Shared cloud
- Validation & early production
- Fastest start
- Innovation Center entry point
No data leaves your jurisdiction. No black-box AI. No compromises on control. This is sovereignty by design.
Sovereign AI infrastructure at scale.
Distributed GPU compute, Kubernetes orchestration, and vector infrastructure — deployed as a private AI cloud with European residency options.
GPU Clusters
Datacenter and on-premises inference for training and production workloads, with tenant-isolated compute boundaries.
Kubernetes Platform
GitOps-managed services, Helm deployments, Firecracker sandboxes, and horizontal agent runtime scaling.
Distributed AI Compute
Horizontally scaled inference, model routing through the AI gateway, and workload-aware scheduling.
Vector Infrastructure
Embeddings pipelines, Qdrant vector stores, and governed retrieval integrated with the Data Layer.
AI Datacenter
European deployment options with residency controls, procurement-ready documentation, and sovereign hosting patterns.
Private AI Cloud
Fully private or air-gapped environments for regulated sectors — no dependency on external inference APIs.
Frontier reasoning and a Dutch language foundation.
Supernova pushes reasoning at the frontier. Geitje 2 grounds language, context, and Dutch-market understanding. The platform connects both into governed, production-grade systems.
Supernova
Efficient, verifiable reasoning systems.
Post-transformer architectures focused on efficiency, verifiability, and benchmark-led claims — reasoning you can inspect, evaluate, and deploy under European governance requirements.
- Verifier-based training and neuro-symbolic hybrids
- Benchmark-first: ARC-AGI, theorem proving, planning
- Designed for sovereign deployment
Geitje 2
Built for the Netherlands. Powered by agentic AI.
Our Dutch language foundation — the preprocessor layer that understands Dutch language, local context, and business workflows better than generic models.
- Dutch-first language, idioms, and domain context
- Sits upstream of tools and agents
- Sovereign deployment path — no foreign APIs by default
Deploy on sovereign infrastructure
Talk to our team about running the full platform on infrastructure you control — with governance, audit evidence, and procurement-ready documentation.