Cellule-PRO — ArchitectureENTERPRISE

Self-contained LLM inference microsystem · atom → pool → federation → governance

Showcase · documentation centerpiece

Control room — live cluster topology

Real-time view of every atom pulsing across your federation: which workers are busy, which proxies migrate, which links carry traffic right now. The single page that explains the entire microsystem in motion.

Open the control room →
Private federated cluster · no data leaves the customer infrastructure IT / EMPLOYEES / API admin SPA · GDPR employee portal · Bearer token API internal chat · OpenAI-compatible SDK · CLI · project RAG HTTPS · WSS · Bearer sk-cellule-* · zero runtime Internet dependency POOL (Docker enterprise) FastAPI backend · vector storage · WS push Adaptive routing · long-term memory · OpenAI-compat API Project RAG · replicated model catalog · dynamic balancer Air-gap runtime image · multilingual embedder baked in → layer 2 · the microsystem backend ATOM uses the PCs / servers you already own Win / Mac / Linux · GPU or CPU auto-onboarding · dormant compute → heterogeneous fleet by design FEDERATION mesh of N paired pools Ed25519 · signed replication · RAID symmetric cross-pool admin HA → layer 3 GOVERNANCE & COMPLIANCE Offline JWT Ed25519 license · initialization ceremony GDPR audit trail · access/rectif/erasure rights (articles 15-22) Incidents & Migrations · admin pilot · configurable failsafe Runtime UI flags · zero hidden env var → layer 4 · control + auditability + sovereignty An atom = one machine that computes. Interchangeable, stateless (except keypair). A federation = a cluster of cryptographic trust. Continuous optimization of existing compute · no data leaves the customer infrastructure · no central pool Opt-in safety feature flags · 1-command rollback · single shippable Docker image
Atom (layer 1) — compute
Pool (layer 2) — orchestration
Federation (layer 3) — mesh
Governance (layer 4) — audit

Layer 1 · Atom

  • • Uses your existing fleet: office PCs, workstations, GPU servers
  • • Multi-OS: Windows / macOS / Linux · dormant CPU or GPU
  • • Auto-onboarding: hardware ↔ adapted-model matching
  • • Persistent Ed25519 identity · stateless except identity
  • • Continuous optimization of available compute power

Layer 2 · Pool

  • • FastAPI backend · vector storage · WS push
  • • Adaptive routing · long-term memory · project RAG
  • • Cross-site replicated model catalog · balancer
  • • Air-gap runtime image (embedder baked in, offline)
  • • Modular business plugins (6-pillar microsystem)

Layer 3 · Federation

  • • Mesh of N Ed25519-paired pools (proprietary trust scale)
  • • Signed cross-pool replication (multi-site RAID)
  • • Symmetric cross-pool admin HA
  • • RAG fanout, zero-data-out preserved
  • • Auto-orchestrated workers/proxies (load + failure)

Layer 4 · Governance

  • • Offline JWT Ed25519 license (ceremony)
  • • GDPR articles 15-22 audit (export/rectif/erasure)
  • • Incidents & Migrations UI (configurable failsafe)
  • • Runtime UI flags (zero hidden env var)
  • • Global auto/manual cluster toggle