Neuro-Symbolic AI
Saros is Novatech’s AI engine for improving the performance and reliability of AI systems used in planning and decision-making.
It is designed to help these systems produce better plans, more reliable recommendations and fewer errors requiring human correction—particularly in complex industrial and infrastructure applications.
Saros focuses on three practical improvements in AI applications.
Recommendations that account for relevant evidence, competing objectives and operating constraints.
Fewer outputs that require rejection or correction before they can be used.
Less time spent on repeatable checking, and more time available for analysis and judgement.
The contribution is assessed against the application’s existing approach: whether it produces more useful results, requires less intervention or achieves the required quality at a lower total cost.
The architecture combines three complementary capabilities.
AI models handle tasks that are difficult to express entirely through fixed rules: interpreting documents, identifying patterns, estimating quantities or proposing candidate solutions.
The application defines the concepts, relationships and conditions that matter.
In a measurement system, this can include what each question measures and how an answer contributes to a score. In a planning system, it can include available resources, dependencies, financial limits and operating constraints.
These representations participate in computation. They are not simply background information included in a prompt.
Software evaluates the conditions that have been formalised. Depending on the application, the result can guide further generation, inform planning, reject an invalid proposal or determine a calculation.
These components do not require a single execution pattern. They may interact during planning, operate within an agentic system or support the processing and evaluation of an analytical result.
Consider a purchasing recommendation. It may appear reasonable while exceeding the available budget, ignoring a supplier’s lead time or relying on inconsistent stock records.
Different parts of the system address different failures.
Data processing can reconcile records before analysis. A calculation can establish the cost of a proposed purchase. A constraint can check that cost against the budget. A planner or agent can then use the findings to revise the proposal.
The contribution is not simply producing another answer. It is making relevant information and requirements affect how the result is obtained and assessed.
Performance is assessed against the task and the conditions in which the system will be used.
Relevant measures include:
Passing computational checks does not establish that every conclusion is correct. Evaluation must also examine whether the system’s assumptions, representations and outputs are appropriate for the application.
AI-assisted instrument design and behavioural measurement
FlexPulse-EU connects questionnaire preparation with an explicit behavioural measurement schema.
Models assist with planning and generating survey content. Validation checks examine its relationship to the measurement requirements and return feedback for correction. Human review remains part of preparation.
Once an instrument is published, defined measurement and mapping rules determine how responses are scored, rather than asking an LLM to reinterpret each participant’s answers.
This provides a concrete implementation of the interaction between model-generated content, explicit representations and deterministic processing.
View FlexPulse-EU →Purchasing and inventory decisions
A decision system combining inventory, sales and financial information to prepare purchasing recommendations under long replenishment times, variable demand and working-capital constraints.
Its implementation brings data preparation, analysis, scenario projections and operational requirements into the same decision process. Recommendations are reviewed by the business before action.
View StockGPT Engine →Policy analysis and scenario exploration
A decision harness that organises evidence and explores possible financial, environmental and social consequences of policy options.
Its architecture was subsequently extended for operational and financial applications in StockGPT Engine.
View Blue World →Saros draws on research in neuro-symbolic AI and model-based verification.
LLM-Modulo describes an interaction between model-generated proposals and external checks, with verification results guiding subsequent attempts. Related experimental work evaluates this approach in planning and scheduling.
These publications inform the architecture. The performance of each implementation must be established through application-specific evaluation.
We work with industrial and research partners on AI applications involving planning, analysis and consequential decisions.
The starting point is a specific technical problem: the information available, the conditions the system must respect and the evidence needed to assess its results.