October 2026
Project website online
From now on, we report on the progress of NEVAKI on this page: on interim results, events and publications.
Neuro-symbolic validation architecture for trustworthy AI systems
A research project of Nibiru Systems GmbH and HTW Dresden
Diese Maßnahme wird mitfinanziert durch Steuermittel auf der Grundlage des vom Sächsischen Landtag beschlossenen Haushaltes.
AI results that can be verified in a provable way
NEVAKI is a joint project of Nibiru Systems GmbH and HTW Dresden University of Applied Sciences. It is funded as an InnoTeam under the MINT skilled workers programme of the Free State of Saxony by the European Social Fund Plus (ESF Plus).
AI systems increasingly make decisions in safety-critical and regulated areas, where errors can have serious economic, legal or social consequences. Yet their reliability is still assessed mainly statistically, on test data. Whether a system actually meets domain-specific, legal or safety requirements can hardly be demonstrated this way.
The EU AI Act makes this gap pressing: for high-risk applications it requires demonstrable risk management, validation and technical documentation. Companies therefore have to do more than build powerful AI – they have to show why, and under which conditions, it can be trusted.
Limited generalisation
When a model meets cases at the edge of its training data, it can decide wrongly without test metrics predicting it.
Opaque decisions
What a neural network bases its decision on usually remains unclear – and so does whether it complies with domain or legal requirements.
Hard to verify formally
Formal verification of neural networks has so far only worked for partial aspects or heavily simplified models.
No system perspective
Validation often looks at the model alone, not at its interplay with pre- and post-processing, decision logic and operation.
Understand neurally, verify symbolically
NEVAKI combines the strength of large language models in understanding natural language with the reliability of formal logic. Requirements, rules and domain knowledge are modelled explicitly and linked to the outputs of the AI. This turns a purely statistical assessment into a knowledge- and requirements-based check.
LLM-based semantic parser
Breaks texts down step by step and turns statements into formal structures such as predicates, relations, rules and constraints. Automatic checks for syntax, consistency and completeness catch errors early.
Rule and constraint model
Obligations, prohibitions, conditions and exceptions from compliance requirements, medical guidelines or legal norms become reusable formal rule sets.
Deterministic logic solver
Acts as a hard verification layer and checks consistency, rule compliance and satisfiability. Uncertain or faulty outputs of the language model are detected, discarded or reinterpreted.
Explanation component
Translates proofs, rule violations and conflicts into understandable justifications. This creates a continuous chain of explanation from the input to the decision.
An orchestration layer couples the components: the language model provides interpretations with confidence scores, and the symbolic check secures them. The system does not blindly trust neural predictions but verifies them systematically.
Tested on real documents
An integrated demonstrator tests the architecture on texts and documents from four demanding domains. The goal is to show that it reliably formalises complex domain content, checks it deterministically and explains the results comprehensibly.
Arguments with many references to laws and regulations that have to be coherent.
Precisely worded claims and features with high demands on completeness and consistency.
Internal and external requirements full of obligations, prohibitions, conditions and exceptions.
Specialist texts whose statements have to be consistent with medical guidelines and rules.
Five consecutive work packages lead from the state of the art through architecture and prototypes to a validated demonstrator.
Analysis of existing approaches
Systematic analysis of existing approaches to neuro-symbolic AI and the validation of trustworthy AI – from translating natural language into formal logic to LLM-based checking methods. It draws on literature, patent and technology research as well as interviews with experts from research and industry.
05/2026 – 06/2026 (2 months) · Milestone M1 – Overview of the state of science and technology including gap analysis
Design of the overall concept
Design of the reference architecture and specification of its modules and interfaces: semantic parser, rule and constraint model, logic solver, orchestration and explanation component. First modules are implemented as prototypes, and a reproducible test environment is set up.
07/2026 – 12/2026 (6 months) · Milestones M2 – Project website for science communication; M3 – Technological overall concept of the validation architecture
Prototypes and proof of function
The core components are implemented as working prototypes and coupled. Experiments with text and table data from practical use cases show whether natural language can be formalised reliably and checked reproducibly – including edge cases and faulty inputs.
01/2027 – 09/2027 (9 months) · Milestones M4 – Integrated prototypes of the core components; M5 – Experimental proof of functionality
Validation in practice
An integrated laboratory demonstrator brings all modules together and is tested on documents from law, patents, compliance and medicine – assessed by completeness of the formalisation, rule compliance and quality of the explanations.
10/2027 – 06/2028 (9 months) · Milestones M6 – Integrated laboratory demonstrator in operation; M7 – Technology validated in a relevant environment
Documentation and exploitation
The results are documented and published. In addition, a roadmap for further development is drawn up, and intellectual property, exploitation and possible contributions to standardisation are assessed.
07/2028 – 09/2028 (3 months) · Milestones M8 – Project documentation and scientific publication; M9 – Exploitation and further development concept
Five consecutive work packages lead from the state of the art through architecture and prototypes to a validated demonstrator.
1.
Analysis of existing approaches
Systematic analysis of existing approaches to neuro-symbolic AI and the validation of trustworthy AI – from translating natural language into formal logic to LLM-based checking methods. It draws on literature, patent and technology research as well as interviews with experts from research and industry.
05/2026 – 06/2026 (2 months) · Milestone M1 – Overview of the state of science and technology including gap analysis
2.
Design of the overall concept
Design of the reference architecture and specification of its modules and interfaces: semantic parser, rule and constraint model, logic solver, orchestration and explanation component. First modules are implemented as prototypes, and a reproducible test environment is set up.
07/2026 – 12/2026 (6 months) · Milestones M2 – Project website for science communication; M3 – Technological overall concept of the validation architecture
3.
Prototypes and proof of function
The core components are implemented as working prototypes and coupled. Experiments with text and table data from practical use cases show whether natural language can be formalised reliably and checked reproducibly – including edge cases and faulty inputs.
01/2027 – 09/2027 (9 months) · Milestones M4 – Integrated prototypes of the core components; M5 – Experimental proof of functionality
4.
Validation in practice
An integrated laboratory demonstrator brings all modules together and is tested on documents from law, patents, compliance and medicine – assessed by completeness of the formalisation, rule compliance and quality of the explanations.
10/2027 – 06/2028 (9 months) · Milestones M6 – Integrated laboratory demonstrator in operation; M7 – Technology validated in a relevant environment
5.
Documentation and exploitation
The results are documented and published. In addition, a roadmap for further development is drawn up, and intellectual property, exploitation and possible contributions to standardisation are assessed.
07/2028 – 09/2028 (3 months) · Milestones M8 – Project documentation and scientific publication; M9 – Exploitation and further development concept
What the project will deliver
October 2026
From now on, we report on the progress of NEVAKI on this page: on interim results, events and publications.
We will keep adding talks, workshops and publications from the project here.
Whether AI can be used where reliability counts depends on whether its behaviour can be verified in a provable way.
The EU AI Act requires demonstrable validation for high-risk applications. NEVAKI creates the technical basis for making rule compliance verifiable and documentable.
Traceable, rule-compliant AI lowers liability and investment risks and shortens development and approval processes – especially for small and medium-sized companies.
The InnoTeam builds expertise at the intersection of AI, knowledge representation, software engineering and system assurance – and creates qualified jobs in Saxony.
NEVAKI strengthens Dresden as a centre of expertise for trustworthy AI, embedded in an ecosystem of Silicon Saxony, Smart Systems Hub and technology-driven companies.
An InnoTeam from industry and academia
In the InnoTeam, staff from Nibiru Systems GmbH and HTW Dresden work on the project together. Industrial development experience meets scientific methodology – strengthening knowledge transfer between university and industry.
Coordinator
Founded in Dresden in 2022, the company develops data-driven solutions and AI applications – from data analysis through AI strategies and prototypes to integration into existing IT landscapes. Nibiru coordinates NEVAKI and is responsible for the practical design of the architecture, the development and demonstration environments and the exploitation of the results.
nibiru-systems.comScientific partner · Chair of Database Systems, Prof. Dr.-Ing. Maik Thiele
The chair researches data-driven, AI-based analysis and validation systems, including uncertainty measurement in language models and ethical AI. HTW Dresden is responsible for the scientific and methodological foundation, experiments and evaluation studies, and brings the results into teaching and theses.
htw-dresden.deAre you interested in NEVAKI, would you like to join a workshop or contribute your own use cases as a practice partner?
We look forward to hearing from you.
Are you interested in NEVAKI, would you like to join a workshop or contribute your own use cases as a practice partner?
We look forward to hearing from you.
E-Mail:
Phone:
Co-funded by the European Union and the Free State of Saxony
NEVAKI is co-funded as an InnoTeam under the MINT skilled workers programme of the Free State of Saxony by the European Union through the European Social Fund Plus (ESF Plus). This measure is co-financed with tax funds on the basis of the budget passed by the Saxon state parliament.
The European Social Fund Plus is the European Union's main instrument for investing in people – in employment, education and skills.
Diese Maßnahme wird mitfinanziert durch Steuermittel auf der Grundlage des vom Sächsischen Landtag beschlossenen Haushaltes.
High-quality AI and data science consulting for our customers
Pages
Navigation
Top
Approach
Work plan
Results
Contact
Socials
High-quality AI and data science consulting for our customers
Pages
Navigation
Top
Approach
Work plan
Results
Contact
Socials
© Nibiru Systems GmbH 2026
