// Sovereign AI · v2.2
Platform Architecture
3-server topology with local LLM inference, vector memory, and an audited relational database. All inter-server traffic flows through an encrypted WireGuard tunnel. No data ever leaves EU jurisdiction — with an encrypted disaster-recovery copy replicated nightly to an offsite Storage Box in Helsinki.
Sovereign AI · infrastructure 3-server topology · sovereign LLM
RAG · vector memory · WORM audit
Sovereign · GDPR · NIS2
// Flow legend
HTTPS · external traffic
Front → GPU · WireGuard
GPU ↔ Data · vectors & LLM
Data internal · backup & audit
RAG · top-k retrieval
Offsite DR → Helsinki · nightly
FRONT · Interface
dedicated EU bare-metal · high-core-count CPU
GPU · AI Cognition
dedicated EU bare-metal · single enterprise GPU
DATA · Sovereign Memory
dedicated EU bare-metal · ECC RAM · NVMe + multi-TB cold archive
OFFSITE · Disaster Recovery
Hetzner Storage Box · Helsinki · geo-separated
Internet
external traffic · web · mobile · API clients
Edge Gateway
Nginx · TLS 1.3 · Let's Encrypt · WAF · rate-limit · :443
Next.js · App
SSR · React 18
Main authenticated frontend.
Astro · Portal
static + ISR
Public portal · open data.
FastAPI
microservices · Py 3.12
app-api · portal-api · persona · cognition · bridge.
OCR worker
Tesseract + LLM
OCR of incoming documents.
Enricher
24/7 daemon
Registries · public sources · scraping.
Nginx
reverse proxy
TLS termination · routing · mTLS.
:8001
vLLM · Qwen 3-8B
AWQ 4-bit · 32k ctx
Sovereign LLM · Apache 2.0 · multilingual.
:9879
BGE-M3
1024-dim multilingual
Embeddings for RAG and semantic search.
:8765
LoRA fine-tune
rank 16 · α 32
Domain adapters · fine-tuning on own Q&A pairs · every run tracked (params, metrics, artifacts).
:8099-101
Reranker pool
bge-reranker-large
3 instances · top-50 → top-5.
Agent Swarm CC1–CC7
parallel agents
OCR triage · enrichment · controls · self-learning.
Bridge API
FastAPI · cognition
Self-learning
Q&A → adapter
:4222
NATS · event bus
heartbeat + dead-man's-switch
:5432 · :6432
PostgreSQL 18 · PgBouncer
62 schemas · deterministic entity resolution · RLS multi-tenant
domain
entities
regulatory
knowledge
content
ai_state
persona
provenance
public
:6333
Qdrant
vector database
Domain-adapted collections: legal, knowledge, entities, procurement, press, court records.
1024-dim · cosine sim.
:7687
Neo4j
graph · Cypher
Entity hierarchies · forensic relationship detection.
Ownership edges are bi-temporal (valid-from/valid-to), rolling out to other edge types.
:6379
Redis
cache · queue
Sessions · rate-limit · jobs.
:9000
MinIO · S3-compat
object storage
Documents · images · PDF · scanned materials.
Write-guard triggers
BEFORE INSERT/UPDATE, targeted
Deployed where correctness is non-negotiable — brand normalization, content-safety guards. Extending to full field-level history is scoped, not blanket-claimed.
Cold backup archive · local
daily + monthly full dumps · continuous WAL archiving
Verified RPO ≤ 24h on the primary database (daily automated dump, confirmed current) plus a separate nightly job replicating to the offsite Helsinki box below. Completeness audits run periodically and are allowed to fail loudly — the last one flagged two components needing attention, which is the point of running it.
ssh :23
Storage Box · Helsinki
Hetzner offsite · geo-separated · nightly automated
Independent nightly jobs from both compute nodes (GPU 20:00, DATA 21:00) — database dumps, LoRA adapters, configs. Both confirmed successful on the most recent run. True 3-2-1 — survives full datacenter loss.
⟦ WireGuard VPN · encrypted mesh ⟧ ⟦ WireGuard VPN · :5432 · :6333 · :7687 · :6379 · :9000 ⟧ ⟦ encrypted rsync · SSH :23 → Helsinki ⟧ HTTPS · :443 TLS 1.3

// Agent governance — candidate until evidence promotes it

Every agent action starts as an untrusted candidate. It earns the right to execute through scope, a tool gate, evidence, human oversight, and a scoped time-bound runtime permit. Automation expands in proportion to reversibility — speed is earned, not assumed. Each action is recorded to a tamper-evident audit trail.

1
Candidate Signal
plan/step proposed
2
Passport Scope
owner + plan identity
3
Tool Gate
4-level permission
4
Toll Evidence
context + RAG grounding
5
Human Review
approval pause
6
Runtime Permit
scoped · TTL · uses
7
Broker Receipt
execution logged
8
Evidence Trail
audit, tamper-evident
Runtime permit + TTL

Each action receives a scoped, time-bound permit before execution: read-only 1h / 100 uses, mutating 5 min / 3 uses. A mutating action without a valid permit is blocked. Permit lifecycle (issued · consumed · expired) is recorded for replay and accountability.

Graduated reversibility

A tool earns a trust tier from its track record. Proven tools receive longer, larger permits; autonomous plans may use proven, reversible mutating tools once an operator marks them eligible. Permission ceiling: read-onlymutateshuman approvalforbidden.

// Network & security

WireGuard VPN between all three servers, TLS 1.3 for external traffic. External access exclusively through Front. Encrypted offsite disaster-recovery replication to a geo-separated Storage Box in Helsinki (Finland) — true 3-2-1.

// Legal framework & compliance

The platform is built in accordance with Croatian and EU regulatory frameworks:

Public Procurement Act OG 120/16 GDPR EU 2016/679 Right to Access Information Act OG 25/13 NIS2 EU 2022/2555 Local & Regional Self-Government Financing Act OG 127/17 GDPR Implementation Act OG 42/18

// Sovereign principles

  1. Data never leaves EU jurisdiction.
  2. Open-source model (Apache 2.0) — no vendor lock-in.
  3. Every AI response is auditable and reconstructible.
  4. The client controls what the model learns and what it discards.