# Congruence

> Congruence is a FER Zagreb spin-off building sovereign AI infrastructure for regulated industries across Europe. The company operates two lines: an open R&D line contributing five IPCEI AI primitives to the European AI ecosystem, and a commercial line deploying production sovereign AI systems — starting with Congruence Sovereign, a 24-module supervisory intelligence platform for Basel III and Solvency II financial regulation. On-premise, air-gapped, zero cloud dependencies.

- Founded: 2026, Zagreb, Croatia
- Type: FER Zagreb spin-off (Faculty of Electrical Engineering and Computing, University of Zagreb)
- Domain: Sovereign AI, financial regulation, EU AI Act compliance
- Status: IPCEI AI direct partner, Thales Group collaboration (CRT x STRATOS²-AI)
- Contact: info@congruence.hr

---

## Home

URL: https://congruence.hr/

### Mission

Congruence builds open R&D infrastructure for the European AI ecosystem and production-grade sovereign AI systems for regulated industries — where AI must be certified, auditable, and independently controlled.

AI that belongs to the institution running it.

Most organizations will never own their AI. That is a structural problem. Congruence exists to close that gap — with precision.

### Two Lines Architecture

Congruence operates two complementary tracks. The R&D line advances European AI capability through open contributions — primitives that any institution can adopt. The commercial line deploys those primitives as production systems that institutions operate inside their own infrastructure.

#### R&D Line — Open research primitives for the European AI ecosystem

Five foundational R&D contributions developed under IPCEI AI — solving the hard infrastructure problems of sovereign, certified, and federated AI deployment. Released as open-source building blocks that any European institution can use.

1. Federated RAPTOR-GraphRAG (F-RG²) — Cross-border retrieval (Open source)
2. Sovereign Mixture-of-Experts (SS-MoE) — Federated models
3. Speculative Decoding for Sovereign Inference — Latency parity
4. Conformal Prediction — EU AI Act certified uncertainty (Provable guarantees)
5. Neurosymbolic Reasoning (NRRE) — Regulation-as-code

#### Commercial Line — Production sovereign AI systems for regulated industries

Purpose-built AI systems that regulated institutions deploy on infrastructure they control.

1. Congruence Sovereign — supervisory intelligence platform (Live, 24 modules)
2. Sector-specific foundation models (regulatory data) — In development
3. Open EU AI Act compliance framework — Apache 2.0
4. Sovereign AI Factory — Tier 4 GPU infrastructure (SAF-Croatia)
5. AIaaS platform for European regulated sectors — 2026-2031

### EU AI Positioning

#### IPCEI AI — Direct Partner
Congruence participates as a direct partner in IPCEI AI, contributing five state-of-the-art R&D work packages with demonstrable spillover across 27 Member States and every regulated sector. Not a beneficiary — a contributor to common European interest.

#### EU AI Act — First Mover
The EU AI Act is fully enforceable in 2026. Every high-risk AI system in Europe needs conformity assessment. No MLOps tool currently provides AI Act integration. Congruence is building the missing compliance layer — and releasing it open-source under Apache 2.0.

#### Sovereign AI Factory — Croatia
EUR 17M investment. 1 MW Tier 4 GPU facility in Croatia, scalable to 5 MW. Federated learning aggregation on auditable sovereign hardware. The physical infrastructure that makes European AI sovereignty operational — not theoretical.

### Founders Summary

- Luka Petrovic, Ph.D. — Co-Founder, Assistant Professor, FER Zagreb. PhD in Robotics (2022). Research: high-dimensional trajectory optimization, Gaussian processes, neurosymbolic AI. Work package leader on EU Horizon Europe projects. Visiting researcher at Oxford Robotics Institute and KIT.
- Vlaho-Josip Stironja — Co-Founder, PhD Researcher, FER Zagreb. Research: robust multisensor mobile robot localization, ego-motion estimation. Google DeepMind ML Summer School participant. Expertise in LLM workflow engineering, estimation theory, sensor fusion.

---

## Platform — Congruence Sovereign

URL: https://congruence.hr/platform.html

### Overview

Congruence Sovereign is a supervisory intelligence platform for financial regulation. Five architectural layers. 24 production modules. On-premise, air-gapped. Full LLM audit trail on every AI decision. Human-in-the-loop at every lifecycle transition. Basel III and Solvency II — natively — on infrastructure the institution controls, with zero cloud dependencies.

Key metrics:
- 5 architectural layers
- 24 production modules
- 2 regulatory frameworks
- 0 cloud dependencies

### Five-Layer Architecture (24 Modules)

#### Layer 01 — Ingest: Data Acquisition (5 modules)
1. **Contextual Data Gate** — Magic-byte validation, metadata envelope tagging, framework-specific routing
2. **Advanced Document Parser** — Multi-strategy PDF/OCR, semantic chunking with overlap, KPI tag assignment
3. **Document Slot Manager** — Framework-inherited slots, completeness tracking, pipeline-type routing
4. **Basel III Quantitative Extractor** — 100+ fields: CET1/AT1/T2, RWA, LCR, NSFR, leverage ratio
5. **Solvency II QRT Extractor** — SCR components, own funds tiering, technical provisions, MCR

#### Layer 02 — Compute: AI Inference (6 modules)
1. **Cognitive Inference Engine** — 24 risk services; up to 5-step chained LLM workflows per calculation
2. **Basel III Risk Engine** — 6 risk types; IR + QRM; importance weights 25/20/20/15/10/10
3. **Solvency II Risk Engine** — 6 risk types via adapter pattern; QRT quantitative supplement
4. **Net Risk Calculator** — 5x5 IR x QRM matrix; composite scoring; intervention ladder mapping
5. **Assessment Executor** — Parallel job orchestration; SSE progress streaming; state machine
6. **BPMN Workflow Engine** — BPMN 2.0 execution; external task API; human-in-loop steps

#### Layer 03 — Memory: Retrieval & State (4 modules)
1. **Knowledge Retrieval (RAG)** — ChromaDB; framework-scoped collections; regulatory knowledge base
2. **Prompt Registry** — Database-backed templates; version history; admin UI editing
3. **Scoring Template Store** — Versioned IR/QRM scales; 5x5 net risk matrix; intervention ladder
4. **Institution & Assessment Registry** — Full lifecycle management; multi-framework assessment support

#### Layer 04 — Govern: Policy & Control (4 modules)
1. **Agentic Guardrails** — Analyst -> Critic -> Feedback loop; adversarial multi-round refinement
2. **Supervisory Review Board** — Multi-assessor scoring; LLM auto-score with human override
3. **RBAC & Authentication** — JWT; Admin / Editor / Viewer; bcrypt; route-level enforcement
4. **Framework Hierarchy Engine** — Self-referential tree; inheritance resolution; jurisdiction tagging

#### Layer 05 — Observe: Audit & Transparency (5 modules)
1. **Glass Box Audit Logger** — Full LLM token trail; prompt version FK per calculation; correlation IDs
2. **Real-Time Progress Streaming** — SSE per-step labels; parent-child job hierarchy; cancellation
3. **Assessment Reporting Engine** — PDF report generation; SAFR XML export; structured supervisory records
4. **CBTT RBS Dashboard** — Purpose-built for Central Bank of Trinidad & Tobago methodology
5. **AZN ORSA Dashboard** — Purpose-built for Slovenian Insurance Supervision Agency (AZN)

### Regulatory Framework Support

#### Basel III — Banking: ICAAP Risk-Based Supervision
Six risk types across IR and QRM calculation methods. Quantitative supplements from regulatory returns — CET1, RWA, LCR, NSFR, leverage. Weighted composite scoring maps directly to supervisory intervention levels.

- Credit Risk (IR + QRM — 5-step LLM workflow): 25%
- Market Risk: 20%
- Liquidity Risk: 20%
- Operational Risk: 15%
- IRRBB (4-step metric extraction): 10%
- Market Conduct Risk: 10%

#### Solvency II — Insurance: ORSA Supervisory Assessment
Six risk types with QRT data import — SCR components, own funds tiering, technical provisions, MCR. Single-entity and group-level submissions.

- Insurance Risk: ORSA adequacy
- Underwriting Risk: QRT supplement
- Market Risk: SCR components
- Credit / Counterparty Risk: Own funds tiering
- Operational Risk: SCR/MCR tracking
- Liquidity Risk: Technical provisions

### Deployment Models

1. **Air-Gapped On-Premise** — Complete network isolation. NVIDIA GPU hardware (A100/L40S/H200). Zero network egress. For central banks and national regulators where data cannot leave the premises.
2. **Private Cloud** — Single-container Docker architecture on AWS GovCloud, Azure Government, or private Kubernetes. Data boundaries at container level. For institutions with cloud mandates requiring sovereign data residency.
3. **Managed SaaS** — Cloud LLM alternatives with per-organization database schemas and configurable data residency. Upgrade path to on-premise at any point.

### Technology Stack
- FastAPI — Async Python API framework. All 24 risk services, assessment orchestration, audit endpoints.
- React 19 + TypeScript — Production frontend. Dashboard, assessment workflow UI, admin panels.
- PostgreSQL 16 — Primary datastore. Institution registry, assessment history, prompt versions, audit logs. Row-level security.
- ChromaDB — Vector database for RAG pipeline. Framework-scoped collections.
- Ollama — Local LLM inference runtime. Model management, GPU allocation. Air-gap compatible.
- Nemotron-3-Nano 30B — Production language model. 30 billion parameters on NVIDIA GPU hardware the institution owns.

---

## Research — Five IPCEI AI R&D Contributions

URL: https://congruence.hr/research.html

Each contribution addresses a foundational technical barrier that prevents European regulated industries from operating sovereign, certified AI at scale. Each produces open-source artifacts. Each generates spillover that extends beyond Croatia and France — to every Member State and every regulated sector.

### 01. Federated RAPTOR-GraphRAG (F-RG²)

Tag: Open source

Hierarchical document retrieval across 13+ Member State regulatory corpora — without exposing source documents. Cross-border regulatory intelligence without centralization.

A federated retrieval-augmented generation system that combines RAPTOR tree-based summarization with GraphRAG knowledge graph extraction. Each participating institution maintains its own document corpus. The federated layer enables cross-border queries — a Slovenian supervisor can retrieve relevant Croatian or French regulatory interpretations without either party exposing underlying documents. The system produces hierarchical summaries at multiple abstraction levels, preserving regulatory nuance that flat retrieval systems destroy.

- Cross-Border Spillover: Regulatory corpora from 13+ Member States become queryable through federated retrieval without data centralization. Any national supervisor gains access to interpretive patterns across jurisdictions.
- Cross-Sector Spillover: Architecture applies to healthcare (clinical guidelines), legal (multi-jurisdiction case law), and environmental regulation (emissions reporting).
- Cross-Actor Spillover: Released as open-source. Any European SME, university, or public institution can deploy on their own infrastructure.

### 02. Sovereign Mixture-of-Experts (SS-MoE)

Tag: Federated

Federated expert model training across IPCEI partners. Each Member State trains on local data; the combined model serves any domain, any jurisdiction — with conformal routing guarantees.

A mixture-of-experts architecture where individual expert modules are trained independently by different IPCEI partners on their local data. A sovereign router — itself trained with conformal prediction guarantees — directs each inference request to the appropriate expert without exposing training data across organizational boundaries.

- Cross-Border Spillover: Each Member State contributes domain expertise through locally trained expert modules. Combined model achieves performance no single-nation dataset could produce.
- Cross-Sector Spillover: MoE routing architecture is domain-independent. Financial regulation trains first expert cohort; healthcare, energy, public administration can add sector-specific experts.
- Cross-Actor Spillover: Any institution can train and contribute an expert module. Barrier to entry is a local dataset and GPU allocation.

### 03. Speculative Decoding for Sovereign Inference

Tag: Infrastructure

2-3x inference speedup on sovereign European infrastructure — eliminating the latency penalty that drives enterprises toward US hyperscalers.

A speculative decoding implementation optimized for the sovereign deployment constraint: limited GPU hardware, no cloud burst capacity, strict latency requirements. Uses a small draft model to generate candidate token sequences that the large target model verifies in parallel — achieving 2-3x wall-clock speedup without any change to output distribution.

- Cross-Border Spillover: Every European institution running sovereign AI faces the same latency penalty. Technique is hardware-dependent, not jurisdiction-dependent.
- Cross-Sector Spillover: Any sector deploying LLMs on-premise benefits. Model-agnostic and domain-agnostic.
- Cross-Actor Spillover: Open-source inference optimization library for AI Factories, university clusters, and SME GPU nodes.

### 04. Conformal Prediction for EU AI Act Compliance

Tag: EU AI Act

Distribution-free coverage guarantees for any AI model in any domain. The only technique satisfying EU AI Act Article 15 accuracy requirements with mathematical proof.

A conformal prediction toolkit that wraps any AI model — classification, regression, or generative — with distribution-free uncertainty quantification. Output is not a point estimate but a prediction set with guaranteed coverage rate. The only known technique satisfying EU AI Act Article 15 requirements for accuracy, robustness, and cybersecurity with provable bounds rather than empirical benchmarks.

- Cross-Border Spillover: EU AI Act conformity assessment applies uniformly across all 27 Member States. One implementation, 27 beneficiaries.
- Cross-Sector Spillover: Model-agnostic and domain-agnostic. Financial regulation, medical diagnosis, autonomous systems, environmental monitoring.
- Cross-Actor Spillover: Open-source universal toolkit. Fills structural gap in European AI compliance ecosystem.

### 05. Neurosymbolic Regulatory Reasoning Engine (NRRE)

Tag: Certifiable

EU regulation encoded as machine-verifiable symbolic rules. AI systems that produce certifiable reasoning traces — not just answers.

A neurosymbolic architecture combining neural language model capabilities with formal symbolic reasoning. Regulatory text encoded as machine-verifiable rules. Generates formal reasoning traces independently verifiable against symbolic rule sets. Output: "the following logical derivation, traceable to Article X, Section Y, produces this conclusion." Compliance determinations auditable by design, not by assertion.

- Cross-Border Spillover: Symbolic encoding of Basel III developed for one jurisdiction is directly reusable by any supervisor applying the same framework.
- Cross-Sector Spillover: Insurance (Solvency II), energy (REMIT), environmental (EU taxonomy), data protection (GDPR) all benefit from formal symbolic encoding.
- Cross-Actor Spillover: Symbolic rule sets released as open-source regulatory knowledge artifacts.

### Compound Effects

The five contributions are designed to compound:

1. **MoE + Conformal Prediction**: Sovereign MoE router uses conformal prediction to guarantee expert selection accuracy. Certified expert routing, not probabilistic guessing.
2. **RAPTOR-GraphRAG + Neurosymbolic**: Federated document retrieval feeds the neurosymbolic reasoning engine. Two layers of auditability.
3. **Speculative Decoding + MoE + Edge Slicing**: Combined effect — sovereign inference performance matching cloud alternatives on institutional hardware.

---

## Founders

URL: https://congruence.hr/founders.html

Congruence is a FER Zagreb spin-off — two researchers who have built real AI systems at production scale and encountered, at each stage, the infrastructure problems no existing provider had solved. The intellectual content is not derived from abstraction. It is the product of operating real systems inside real institutions.

### Luka Petrovic, Ph.D.

Co-Founder. Assistant Professor, Faculty of Electrical Engineering and Computing, University of Zagreb.

PhD in Robotics, FER Zagreb 2022. Thesis: "High-Dimensional Trajectory Optimization for Robot Motion Planning Based on Gaussian Processes." Graduated with the highest honors. Research covers motion planning, trajectory optimization, Gaussian processes, neurosymbolic AI architectures, and human-robot collaborative systems.

**Research Leadership:**
- Work package leader: HORACE (Weave/HRZZ/SNSF/ARIS, 2024-2028)
- Work package leader: DRUMS (Erasmus+, 2024-2026)
- Researcher: FITNESS (Horizon Europe EIC Pathfinder, 2023-2027)
- Researcher: AIFORS (Horizon 2020, 2021-2026)

**Industry:**
- Jacobi Robotics Ltd (Silicon Valley, 2024-2026) — motion planning and AI algorithms, Fortune 500 production deployment

**Visiting Research:**
- Oxford Robotics Institute — Dynamic Robot Systems Group, Prof. Ioannis Havoutis
- Karlsruhe Institute of Technology — IPR Lab, Prof. Bjorn Hein

**Awards:**
- Award for Young Scientists (University of Zagreb 2023)
- Silver Plaque "Josip Loncar" (outstanding PhD thesis 2022)

**Selected Publications:**
- IEEE T-SMC (2022, 2026)
- IEEE T-RO (2021)
- Robotics and Autonomous Systems (2020, 2021, 2023, 2024)
- ICCV 2025
- IROS 2019/2025

**Service:**
- Section Editor, Automatika (Taylor & Francis)
- Associate Editor, IFAC World Congress Busan 2026
- Program Chair, ROSCon Croatia 2026

Relevance to Congruence: PhD work on Gaussian process-based motion planning directly informs the probabilistic inference architecture in Congruence Sovereign's Cognitive Inference Engine. Neurosymbolic architectures research underpins the NRRE contribution (R&D line 05).

### Vlaho-Josip Stironja

Co-Founder. PhD Researcher, Faculty of Electrical Engineering and Computing, University of Zagreb.

PhD candidate researching robust multisensor mobile robot localization. Research areas: ego-motion and ego-velocity estimation (radar, IMU, LiDAR, camera, GNSS), online sensor extrinsic and temporal calibration, novel electromagnetic sensor modalities.

**Selected Publications:**
- IROS 2025
- Robotics and Autonomous Systems (MOVRO2)
- MFI 2024 (RAVE framework)

**Awards & Recognition:**
- Best Student Paper Award finalist at IAS-18 and RAAD 2024
- Google DeepMind Machine Learning Summer School — 300 participants selected from 1,700 applicants

**Technical Expertise:**
- LLM workflow engineering
- Estimation theory
- Sensor fusion
- Real-time inference optimization

**Service:**
- Co-organizer, Croatian Robotics Conference 2025

Relevance to Congruence: Sensor fusion and estimation theory expertise directly underpins the real-time sovereign AI inference architecture. LLM workflow engineering experience is the technical foundation of the Cognitive Inference Engine and Agentic Guardrails module.

### Company Context

FER Zagreb spin-off — Faculty of Electrical Engineering and Computing, University of Zagreb.

- IPCEI AI invited participation — EU AI matchmaking, Berlin, March 2026
- Collaboration: Thales Group — CRT x STRATOS2-AI framework (five beyond-state-of-the-art R&D work packages)
- Active Horizon Europe participation across multiple work packages

Congruence operates at the intersection of frontier research and production deployment. The company is structured to contribute open R&D infrastructure to the European AI ecosystem while deploying production sovereign AI systems for regulated industries. This dual structure — open R&D line and commercial line — is not a compromise. It is the architecture that makes both sustainable.