The Center for Data Innovation recently spoke with Aga Kopytko, co-founder of Smabbler, a Poland-based company developing a platform that helps large language models (LLMs) give more accurate answers and avoid generating false information. Kopytko explained how this allows organizations to pull reliable and traceable insights from large collections of documents and data, improving decision making in fields such as life sciences and manufacturing.
David Kertai: What inspired Smabbler’s creation?
Aga Kopytko: Many organizations sit on vast collections of information in the form of documents, reports, emails, and research papersWhen they try to use large language models to make sense of this material, they often run into a core limitation: language models generate fluent text, but they do not build a dependable understanding of the underlying facts.
Smabbler fills that gap by turning unstructured information into a structured, navigable knowledge base. Teams upload their documents, and the platform organizes the content, identifies how ideas connect, and builds a verified map of facts and relationships. Researchers, analysts, compliance teams, and engineers can use this structured knowledge to search, analyze, and make decisions with far more confidence than they could with text generation tools alone.
Inside the platform, Galaxia provides the reasoning layer. It examines the structured facts that Smabbler creates and applies logic to connect them, check consistency, and explain how conclusions are reached. In short, Smabbler organizes the knowledge, and Galaxia thinks with it. Together, they help organizations move from AI that guesses to AI that can explain where its answers come from, supporting work in research, compliance, engineering, and other fields where accuracy and traceability matter
Kertai: What limitations in current LLMs is Galaxia designed to improve?
Kopytko: First, LLMs can hallucinate, meaning they can generate statements that sound plausible but are factually wrong, because they choose words based on patterns in their training data rather than checking those words against verified facts or logical rules. To counter this, the platform adds a layer of deterministic, rule‑based reasoning that verifies information against a structured knowledge graph, helping ensure conclusions are grounded in actual facts rather than statistical guesses.
Second, LLMs operate within a limited context window, which restricts how much information they can consider at once. When prompts grow too large, earlier details fall out of scope, and the model cannot maintain continuity across long documents or datasets. Galaxia addresses this by storing information in long‑term memory structures, allowing it to preserve context over time and reason across large collections of data without losing track of earlier details.
Third, retrieval-augmented generation systems—AI models that combine search with language generation—can attach documents to prompts, but they often do not truly connect the meaning across those documents. They retrieve text, but they do not build a structured representation of how ideas relate. By applying symbolic logic to map relationships between concepts, Galaxia creates an organized, interconnected model of knowledge that supports deeper, more accurate reasoning.
Kertai: How does Galaxia enable models to reason differently?
Kopytko: Most language models generate text one prompt at a time and optimize for linguistic plausibility. They do not create a lasting representation of knowledge. Once a session ends, the structure disappears. Galaxia builds a persistent knowledge model from the documents and data it processes. Instead of predicting likely words, it maps concepts and relationships over time. It then applies symbolic reasoning, using explicit logical rules to infer new relationships from existing knowledge.
Smabbler supports AI models by grounding their outputs in structured reasoning. It’s not a chatbot; instead, teams use it behind the scenes to link AI‑generated responses to verified knowledge and source documents. This lets organizations validate, audit, and confidently use AI in real workflows. With advances in hybrid AI architectures and scalable infrastructure, the platform can process thousands of pages and datasets quickly while maintaining consistency.
Kertai: What are semantic hypergraphs, and what role do they play in your platform?
Kopytko: A semantic hypergraph is a way of representing complex relationships. Unlike a standard graph that links two items at a time, a hypergraph can connect several elements within a single relationship. That makes it useful for modeling things like scientific claims, multi‑step processes, or rules with multiple conditions.
Galaxia uses the semantic hypergraph as its long‑term memory. It organizes information across documents, resolves ambiguous terms, preserves context, and tracks the source of each claim. This connected structure lets the system reason across data rather than store isolated facts.
Kertai: Can you share any real-world examples of Smabbler in use?
Kopytko: In life sciences, the platform analyzed thousands of pages of clinical research and mapped relationships between medications and common side effects. By structuring this information into connected knowledge graphs, it helped researchers identify patterns and gaps more efficiently.
In financial services, Smabbler converted dense regulatory text into a structured compliance model with clear traceability to source documents. This approach reduced review time, improved audit readiness, and allowed teams to focus on higher-level analysis rather than manual verification.
In manufacturing and product design, the platform unified engineering notes, specifications, pricing data, and requirements into a single connected system. Teams detected inconsistencies earlier, understood cross-dependencies more clearly, and made decisions with greater confidence because they could trace each conclusion back to its originating data.
