● AI OS — active 24/7RINET.DEV — CIVIC INTELLIGENCE OS ↗ RI.NET DEV → CIVIC INTELLIGENCE OS

The Cognitive Operating System That Builds Itself other AI systems.

RiNET is not an application. RiNET is an operating system — modular, autonomous, predictive. One brain. Six platforms. 32M+ structured records. Zero downtime. Sovereign EU infrastructure. Zero cloud dependency for inference. While the industry POCs LangChain demos, we operate.

Launch demo Explore the platforms
// Topology

Three servers. One mind.

EU-sovereign, three-node cognitive mesh. WireGuard over 10.10.0.0/24. Cognition on dedicated GPU, memory on the I/O node, public surface at the edge — every layer split for blast-radius isolation.

wg0 · 10.10.0.0/24 · WireGuard mesh GPU 144.76.68.5 cognition DATA 49.12.151.125 memory FRONT 195.201.109.212 edge
GPUGEX44
vLLM · Ollama (qwen3:14b, llama3.1:8b) · BGE-M3 reranker · 60+ rinet-* services · 24×7 enrichment.
DATAAX102-U
PostgreSQL 18 · Qdrant (54 collections, 18.9M vectors) · Neo4j 5 · MinIO · Redis · WORM-retained backups.
FRONTAX42-U
nginx + Authelia · cortex-sport (Next.js) · klasik SPA · public surface; zero data egress beyond signed responses.
9.2M+Total records
5.6M+Knowledge facts
18.9M+Vectors in brain
449KCompanies analyzed
320KPersons in DB
612AI agents active
1.6K/hLearning — non-stop
// Platforms

Six platforms. One mind.

Every platform shares the same cognitive core — what one learns, all know. What one ingests, all access. Cross-platform knowledge propagation in real time, anchored to a sovereign knowledge graph.

🏛️
Civic Intelligence
Cross-jurisdictional public-sector BI. Budgets, procurement, corruption signals, anomaly detection, entity-level forensics. 338K Croatian companies, 75K procurement contracts, 190K persons — all linked, all transparent. The only platform indexing every Croatian limited company cross-referenced with every EOJN procurement contract, every FINA financial filing, every Narodne novine legal change.
rinet.dev — Civic Intelligence OS → app.rinet.one →
🇪🇺
EU Funds Assistant
AI expert for EU funding calls. 32K laws, 309K legal articles, hybrid semantic + reranker retrieval. Drafts documentation, scores success probability, prepares submissions. Production-grade. No comparable system.
eu.rinet.one →
📰
AI Portal
Autonomous technology editorial portal. A digital persona writes, publishes, comments — 100% autonomously. House + Nicholson + Bourdain editorial voice. Showcase of RiNET autonomous content capabilities.
portal.rinet.one →
👤
Digital Persona
Your digital twin — remembers, learns, grows. Digital legacy vault. 19 languages, GDPR, AES-256 encryption, post-quantum migration in progress. Remain present after death.
me.rinet.dev →
📚
TNT Reading Room
Alan Ford digital collection — 500+ comics indexed, AI-powered semantic search, OCR pipeline, character knowledge graph. For the legendary TNT group.
tnt.rinet.dev →
DEV
🧠
DABI AI Suite
Umbrella for all AI modules: Lawyer, IT specialist, MarTech, personas. Every module operates standalone or inside any platform. One API. Infinite composition. — under active development.
rinet.dev →
// AI Brain

The system that builds itself.

700+ verified facts ingested per hour. 5 AI models orchestrated in parallel. GPU saturated 24/7 — by design. Knowledge growing exponentially. Smarter every day than the day before. If you are still wiring Supabase + LangChain + a hosted vector DB, you are already obsolete.

  • LLM waterfall: Groq → DeepSeek → Ollama-local
  • Self-learning: 26,000+/day, zero skip rate, evidence-anchored
  • Forensic engine: 20,360+ findings, entity risk profiling
  • Neo4j knowledge graph: 111K+ nodes, entity risk scoring across companies and persons, temporal relationships
  • 14 specialized AI persona agents — legal, financial, journalist, prosecutor, auditor, social-media analyst, data scientist...
  • DABI chat: 30+ patterns, 100% DB-grounded, zero hallucinations
  • Redis cache: 24-56× faster on heavy queries, hot-path sub-second
  • EVERYTHING passing through the system stays in the brain. Forever. WORM-retained. Tamper-proof.
$ rinet status --live

CORE.............. ONLINE
// AI/ML Stack — Inference & Training

Six models. One GPU. Sub-second response.

vLLM and Ollama running side-by-side on a single RTX 4000 SFF Ada. BGE-M3 embedder over CUDA at 1024-dim. Three reranker instances behind a load-balanced pool. Custom QLoRA fine-tunes shipped to production. A smart router that picks DeepSeek, Groq, vLLM-local, or Ollama per request — not a fixed waterfall, a routing decision. All EU sovereign, all under one shell.

INFERENCE · port 8001

vLLM

PagedAttention serving · OpenAI-compatible

  • Qwen2.5-7B-Instruct-AWQ
  • 4-bit AWQ · half precision
  • 8 192-token context
  • 40% GPU mem · enforce-eager
INFERENCE · port 11434

Ollama

6 local models · hot-swap

  • qwen2.5:7b · 4.4 GB
  • dabi-gemma · 8.0 GB
  • dabi-budget · 4.4 GB · LoRA-merged
  • dabi-3b-hr · 5.8 GB · Croatian
  • gemma3:4b · 3.1 GB
  • nomic-embed-text · 0.3 GB
EMBEDDING · port 9879

BGE-M3 embedder

CUDA · v2.0-fixed · 5+ days uptime

  • 1 024-dim dense vectors
  • Multilingual · 100+ languages
  • Powers 18.9 M vectors / 54 Qdrant collections
  • Self-healing health check daemon
FINE-TUNING · QLoRA

Custom LoRAs

Fine-tunes shipped to production

  • dabi-budget LoRA: 92.7% train / 91.0% eval token accuracy
  • 3 epochs · 783 steps · 1 h 37 m on RTX 4000 SFF Ada (20 GB)
  • dabi-qlora-qwen2.5-3b-v1: HR-domain adapter
  • Daily training timer present, currently paused — reactivate when training data refresh window opens
RERANKING · 8099 / 8100 / 8101

Reranker pool

3 instances · load-balanced

  • Cross-encoder relevance scoring
  • Sub-100 ms typical re-rank on top-50
  • Powers DABI chat + RAG hybrid retrieval
  • Parallel pool · failure tolerance built-in
ROUTING · llm_smart_router

LLM Smart Router

Per-request model selection (not a fixed waterfall)

  • User-facing (brain, chat, app, api) → DeepSeek
  • Editorial (portal, articles) → Groq llama-4-scout
  • Background / bulk (self-learn, ingest) → vLLM-local
  • Failsafe → Ollama dabi-gemma local
// Self-learning

It learns. Continuously. While you read this.

Every minute, dozens of new evidence-anchored facts land in the knowledge graph — extracted from court notices, procurement awards, financial filings, EU register changes, and the press wire. No human in the loop, no model fabrication: every fact carries a source row id and a trust tier. The counter below ticks the live total, drawn from /api/status and interpolated between fetches at the measured rate.

0
knowledge facts (live)
+1,600/h ~38,400/day ● running 24×7
Last 24h — facts ingested per hour peak 2,528 · low 698 · avg ~1,583
−24h−18h−12h−6hnow
// Self-healing

Services that fix themselves.

Ten self-heal daemons watch every other service. When one crashes, the watchdog catches the exit, restarts it inside a back-off envelope, and posts to Telegram if the same service keeps falling. The cognition node alone runs 154 services right now; manual systemctl restart calls are an exception, not a routine.

154
active services on cognition node
10
self-heal daemons watching
24×7
supervisor + watchdog timer cadence
0
unattended manual restarts (last 30d)*
DETECTexit code ≠ 0OOM · OOM-kill · timeoutsystemd · watchdog RESTARTRestart=always · backoff5s · 30s · 5minRestartSec exponential NOTIFYTelegram botif >3 retries / 5min@damir · @oncall recovery confirmed → reset backoff
rinet-supervisor
Master orchestrator. Tracks systemd state across all rinet-* + dabi-* + cortex-* units. Triggers restart when a unit's RestartCount exceeds threshold or its last-output is older than its expected interval.
master-watchdog
Outermost loop — verifies the supervisor itself hasn't deadlocked. Restarts rinet-supervisor if its heartbeat file is stale > 10 min.
rinet-watchdog.timer
Cron-style health probe — pings each critical port, verifies pgbouncer + Qdrant + Neo4j responsive, kicks rinet-self-heal.service on miss.
dabi-health-guardian
DABI sub-orchestrator — reaps zombie persona workers, recycles agent memory caches, restarts dabi-orchestrator-v3 on stuck-state detection.
autofixer
Pattern-matches journalctl for known crash signatures (psycopg2 conn-reset, redis OOM, asyncpg pool exhaustion) and applies the corresponding fix command. Stable since 2026-05-03 retune.
reranker memleak check
Hourly RSS sample on the three reranker workers (8099/8100/8101). If any worker passes its growth threshold, that one instance is bounced — the pool stays warm.
gpu_lock_acquire
Single-writer mutex over the GPU. Long-running training jobs grab a flock; embedders + reranker yield. Prevents two services from blowing up VRAM at the same instant.
autosis-resolver
Sentinel script that walks the agent task queue, kills stuck CC workers (no progress > 30 min), reassigns the task.
guardian-v2
Last-line process watcher — graceful-shutdown handler for cleanups missed by systemd (open transactions, dangling Qdrant snapshots, unflushed Redis writes).
deploy.sh gate
Pre-deploy runs typecheck + service health probe + DB migration dry-run. If any fails, the deploy aborts BEFORE files move — no partial state on disk.
* "0 unattended manual restarts" excludes intentional config changes. The 154-service count and 10-daemon list are real (verified systemctl list-units on the cognition node).
// For developers, regulators, institutions

Ri.NET — Civic Intelligence Operating System

Seven-layer architecture. 600+ autonomous agents. 18.6M vectors across 47 specialized Qdrant collections, including 1.26M court notices and 978k legal documents. Forensic engine surfaces 20,360+ conflict-of-interest signals. Hourly evidence-grounded RAG eval framework. Custom LoRA fine-tunes on dedicated EU GPUs. REST API access. EU sovereign infrastructure. Public documentation.

RINET.DEV
Architecture →
Seven layers. 600+ agents. Vector cortex. Temporal graph. Production.
RINET.DEV/API
REST API →
Three access tiers. Enterprise on request — pricing on inquiry.
RINET.DEV/LEGAL
DPIA + GDPR →
17-part DPIA, GDPR framework, privacy, terms, cookies — all public.
RINET.DEV/DEMO
Schedule demo →
30-minute live system briefing. Real data. No slides. 24h response.

Technical deep-dive. Public documentation. Transparent pricing. Engineering team direct. rinet.dev →

// Modules

Pluggable intelligence.

Every module runs standalone or inside any platform. One API call — instant intelligence.

⚖️
Lawyer AI
32K+ laws. 309K+ legal articles. Hybrid semantic search with reranker. Ask anything about HR / EU legal corpus. Citations included.
💻
IT Geek
DevOps, infrastructure, coding, monitoring. AI that operates servers and builds systems autonomously.
📊
MarTech
Content generation, SEO, social publishing, analytics. Autonomous marketing engine.
🎭
Digital Persona
White-label digital persona. Your style, your voice, your knowledge — in AI form. For brands and individuals.

The future is already operational.
While others talk about AI, Ri.NET is already building it.

If your city, company or project does not yet have an AI that thinks — someone else already does. The gap is widening every hour.

Watch demo →