Most platforms sell you a dashboard.
Ri.NET sells you the ability to see through walls.
A seven-layer forensic architecture — ingestion, normalization, entity resolution, vector cortex, temporal graph, agent swarm, interface. Self-evolving through autogenesis. Blockchain-anchored for regulator-grade audit. Quantum-ready at the cryptographic primitive. six hundred autonomous agents operating a single sovereign instance. One person wrote the Python. One Claude co-wrote the architecture. Thousands of dashboards pretend to be this. Zero are.
Governments generate oceans of data. And drown in them.
In a country with €2.9B in annual public procurement, 32,375 active statutes, 307,329 legal articles, 1.8 million registered companies, and 70,000 civic associations — the question is never whether the data exists. The question is whether anyone can see through it fast enough to matter.
Croatia is our proof point. The same forensic problem exists in every democracy: fragmented registries, inconsistent identifiers, disconnected financial flows, opaque beneficial ownership chains, statutes that reference other statutes that reference court decisions from fifteen years ago. The data is technically public. It is also, in practice, unreachable.
Existing solutions treat symptoms. Dashboards present what you already knew. BI tools require you to know what question to ask. OSINT platforms dump raw feeds and call it transparency. None of them resolve entities across schemas. None of them reason temporally across thirty years of regulatory change. None of them build the graph.
Ri.NET does not compete with these tools. It is the substrate beneath them. The layer that makes "who owns this" and "is this legal" and "has this happened before" answerable in under 200 milliseconds, across every public dataset in a sovereign territory, with citation-grade provenance for every answer returned.
Seven layers. No shortcuts.
Ri.NET is structured as a vertical cortex — seven distinct computational layers, each responsible for one transformation in the path from raw heterogeneous input to structured, query-ready intelligence. Every layer is independently replaceable. No layer assumes anything the layer below cannot prove.
Every layer exposes a stable internal contract. Every transformation is auditable. Every output is traceable to its source. · Read the full architecture documentation →
The Vector Cortex is not a database.
It is a reasoning substrate. Every question asked of Ri.NET descends through the cortex, touches every relevant memory, reconciles against the temporal graph, and returns with a provenance trail. The cortex does not retrieve. It reasons.
Multi-lingual semantic recall
Ri.NET indexes content across seven vector collections, each optimized for a distinct domain: legal text, corporate records, procurement contracts, court rulings, news streams, technical documentation, and conversational memory. Query any language. Retrieve across all.
Entity-aware retrieval
Semantic search is not enough. Ri.NET binds every retrieved passage to the resolved entity it describes. A query about "the procurement officer at KBC Rijeka in Q2 2024" returns the person, their role-at-time, the contracts they signed, the outcomes, and the legal framework in effect that quarter.
Temporal graph reasoning
Every edge carries a validity window. When a company changes ownership, when a statute is amended, when a person leaves a role — the graph does not overwrite. It layers. Ask a question about 2021 and Ri.NET answers with the state of the world in 2021, not the state of the world today.
Citation-grade provenance
Every answer Ri.NET returns carries an audit chain: source document, extraction model, normalization transform, resolution confidence, graph traversal path. There is no "AI said so." There is only "this answer came from these fifteen documents, reconciled against this statute, as amended on this date."
Forensic anomaly detection
Continuous scanning identifies patterns that should not exist. Procurement clusters with identical winning margins. Board members appearing across mutually-exclusive competitor structures. Legal entities with matching fingerprints but different names. Ri.NET finds them. It does not judge them. It surfaces them for the humans who should.
Zero-retention conversational memory
The conversational interface (DABI) is ephemeral by design. Nothing a user asks is retained beyond the session unless explicitly marked for continuity. Sensitive queries leave no trace. Transparent queries build institutional memory. The distinction is user-controlled, cryptographically enforced.
six hundred agents. One orchestrator.
Ri.NET does not run on a cron schedule. It runs on a swarm. A centralized orchestrator decomposes every operational requirement into atomic tasks, delegates to specialized workers, and reconciles results against a shared ontology. The humans who operate the system talk to exactly one agent. That agent does the rest.
The architecture inverts the usual pattern. Traditional platforms expose dashboards and assume humans will compose the right questions. Ri.NET exposes a single reasoning partner (DABI) and assumes humans will ask about outcomes. "Flag every procurement contract above €100K awarded in the last twelve months where the winning bidder was incorporated less than six months before the tender announcement" is not a query to build. It is a sentence to type.
Beneath DABI, six hundred specialized agents execute. One central orchestrator. Five core cognitive domains (Legal, Financial, Civic, Sentinel, Journalist). Sixty mid-level coordinators. Six hundred task-specific workers. The number is not symbolic. It is the measured steady-state fleet size under current operational load.
Agents spawn new agents. Agents retire obsolete agents. The swarm rebalances in real time against workload and observed performance. Every agent executes within strict sandboxing. Every agent logs to the same immutable audit trail. The swarm cannot be observed. The swarm's output can be audited to the character.
The substrate under the substrate.
Numbers are not the product. Numbers are the evidence that the product is real. The following are measured values from production infrastructure as of the current operational quarter.
| Domain | Asset | Volume |
|---|---|---|
| Corporate | Resolved legal entities (companies, institutions, associations) | 1,809,421 |
| Corporate | Beneficial ownership relationships | 447,289 |
| Legal | Statutes, regulations, ordinances (active + historical) | 32,375 |
| Legal | Legal articles indexed and vectorized | 307,329 |
| Legal | Court notices (e-oglasna) with OIB cross-linkage | 92% coverage |
| Financial | RGFI financial statements imported (1996–2025) | 1,864,712 |
| Financial | Pre-2023 records converted HRK → EUR | 100% |
| Procurement | Public tenders indexed (domestic + EU-SIMAP) | 273,118 |
| Procurement | Forensic anomaly findings (critical + high severity) | 1,507 |
| Persons | OIB-enriched individuals with role attribution | 177,041 |
| Asset declarations | Imovinske kartice (public officials) | 161,000+ |
| Vector memory | Indexed passages across seven collections | 6,870,543 |
| Graph | Nodes · Edges | 164,583 · 891,247 |
| Knowledge | DABI Q&A pairs under continuous retrieval | 10,664 |
| Training | LLM fine-tuning records (Croatian legal corpus) | 14,666 |
| Ingestion | Records processed per 24-hour window | ≈180,000 |
| Query latency | Median end-to-end (entity lookup, p50) | 142 ms |
| Query latency | p95 (complex graph traversal) | 487 ms |
| Uptime | Service level (last 90 days) | 99.94% |
| Hallucination rate | Measured across 6,400 evaluator Q&A pairs | 0.3% |
Sovereign-grade, not enterprise-grade.
The difference matters. Enterprise-grade security assumes a commercial adversary. Sovereign-grade security assumes a state-level adversary — or a state-level regulator asking whether your claims are true.
Encryption at rest and in transit
AES-256 for all persistent stores. TLS 1.3 minimum for all transport. Per-tenant key derivation. No shared encryption contexts between clients. Key rotation is automated and auditable.
Blockchain-anchored audit log
Every administrative action — every data ingestion, every configuration change, every user session — is hashed and anchored to a public proof-of-stake chain (Polygon). The audit trail cannot be altered after the fact without global public evidence of alteration.
GDPR-native data model
Data subject rights are not a feature. They are a schema constraint. Right to access, right to rectification, right to erasure, right to restriction — each is implemented as a first-class database operation with guaranteed propagation across all derivative layers.
DPIA-ready by design
A Data Protection Impact Assessment for Ri.NET requires seventeen pages of policy documentation. We wrote them first, then built the system to satisfy them. Every deployment ships with the DPIA package. No ambiguity. No post-hoc justification.
Zero-trust internal fabric
No service inside Ri.NET trusts any other service by default. Every internal call carries a cryptographic assertion of origin, scope, and intent. Lateral movement is architecturally impossible, not merely discouraged.
Forensic readiness
If something goes wrong, the system is optimized for post-mortem. Every operation emits structured telemetry. Every telemetry line is indexed. A forensic investigator can reconstruct any sixty-second window of system behavior to the individual instruction. This is a feature. It is not cheap. It is worth it.
The cortex is addressable.
Ri.NET exposes a structured REST API for entity resolution, graph traversal, semantic search, and forensic pattern queries. Full documentation and live playground access are available to registered developers.
Numbers that matter.
Performance in a civic intelligence context is not about transactions per second. It is about time-to-answer for questions that change policy.
Entity resolution
p50: 42 ms p95: 128 ms p99: 342 ms
Resolves a legal entity across OIB, MB, VAT, historical names, aliases, and mergers. Returns full temporal validity map.
Graph traversal (depth 3)
p50: 186 ms p95: 487 ms p99: 1.2 s
Three-hop relationship search through beneficial ownership, board overlap, procurement participation, and temporal constraints.
Semantic cortex query
p50: 94 ms p95: 223 ms p99: 610 ms
Multi-collection vector retrieval with entity binding, reranking, and citation assembly. RAGAS score: 3.7 / 5.0.
Ingestion throughput
Peak: 14,800 records/min Sustained: 2,100 records/min
End-to-end including normalization, entity resolution, vector embedding, graph insertion, and audit-log anchoring.
A system that improves while you sleep.
Three architectural commitments that distinguish Ri.NET from every platform selling “AI for government.” None of them are marketing claims. Each is implemented in code, measurable in operation, and independently verifiable by any competent third-party auditor.
Self-evolving by design
Ri.NET writes, tests, and merges its own improvements through a disciplined autogenesis pipeline. Dry-run agents propose enhancements, a sandboxed verification harness runs regression tests, binary KPI gates decide whether code advances, and only patches that pass all seven gates are merged. Every merge is signed, every rollback is reversible, every improvement is traceable to its authoring agent and its justifying benchmark. The platform measurably improves between builds without human intervention — and without any of the failure modes that make “self-modifying AI” a horror story in the wrong architecture.
Modular to the core
Every layer, every agent class, every data collector, every embedding backend is independently replaceable. When the current BGE-M3 embedding model is succeeded by a stronger one, Ri.NET swaps the model in-place without a platform rewrite. When a new vector store outperforms Qdrant, it slots in behind the same internal contract. The architecture assumes that every component in the 2026 stack will be obsolete by 2029 — and it is built to absorb that obsolescence without breaking the services built on top of it.
Blockchain-anchored audit trail
Every administrative operation, every data ingestion, every configuration change, every model deployment, every user session — cryptographically hashed and anchored nightly to the Polygon Proof-of-Stake chain. The audit log cannot be altered after the fact without global public evidence of alteration. A forensic investigator five years from now can reconstruct any sixty-second window of system behavior to the individual instruction. This is not a feature we would have built if paying customers had not first asked for it. It is a feature we had already built because we assumed they would.
Quantum-ready cryptographic architecture
The cryptographic layer of Ri.NET is designed for the post-quantum transition. The key-derivation hierarchy accepts pluggable primitives. The audit-anchor signature scheme is replaceable without a data migration. Every long-term cryptographic commitment uses algorithms currently being evaluated for NIST’s post-quantum standardization. When practical quantum hardware becomes commercially available, Ri.NET inherits its cryptographic benefits through an in-place primitive swap, not a platform rewrite. More importantly, the forensic reasoning substrate itself — the entity resolution, graph traversal, and anomaly detection layers — is architected to absorb quantum-accelerated search primitives the moment they become practically deployable on real civic datasets.
Smart learning, not brute-force training
The platform operates a continuous-learning loop rather than periodic retraining batches. Each day it evaluates its own responses against a growing evaluator set of six thousand four hundred Q&A pairs, identifies where confidence disagreements are highest, generates targeted training signal, and incrementally fine-tunes its retrieval and reasoning components. The improvement curve is measured: hallucination rate dropped from 1.2% to 0.3% over 90 days; citation accuracy rose from 78% to 94.1%; RAGAS retrieval score climbed from 1.24 to 3.7 on a 5-point scale. Numbers improve because the system learns from its own mistakes while you read this paragraph.
Sovereign infrastructure, European jurisdiction
Entirely European-hosted. German-regulated physical infrastructure. Dutch-regulated corporate entity. Data never transits outside the EEA under default configuration. No dependency on US cloud providers for core operations. When customers ask “where is my data?” the answer is a specific postal address in a specific country with a specific legal regime — not a Cloudflare shrug.
From transparency OS to decision substrate.
Ri.NET today answers questions. Ri.NET tomorrow answers questions that have not been asked yet. The roadmap is not aspirational. It is sequenced.
Sovereign deployment
First county-level production deployment. Regional regulatory integration. Real-time anomaly alerts to oversight bodies.
Multi-region expansion
Five sovereign territories under live Ri.NET instances. First EU-wide pilot for cross-border procurement monitoring.
EU-wide civic OS
Standardized cross-member-state entity resolution. EU-Lex fully integrated. Forensic pattern sharing across national oversight bodies.
Predictive governance
Policy impact simulation before legislation passes. Early-warning systems for systemic failure patterns. Ri.NET as decision substrate for democratic governments.
The first county is listening.
Ri.NET is currently in pre-deployment with regional authorities, regulatory bodies, and EU-funded governance initiatives. Full case studies will be published as contractual permissions allow.
Regional Government — PGŽ
Coastal-Mountain County pilot for forensic procurement monitoring, regulatory compliance dashboards, and real-time transparency portal for citizens.
EU Funding Advisory
DABI EU Fondovi module in operational use for consultancy workflow automation — matching company profiles to open calls, generating feasibility drafts, drafting legal preambles.
Civic Journalism Network
Continuous anomaly feed for investigative journalists. Pattern-based alerts for procurement irregularities, board-overlap detection, and unusual filing patterns across court registries.
Request a demo or developer access.
Demos are scheduled within 48 hours. Developer access is provisioned automatically after registration. Serious inquiries only.