ETD

• artificial intelligence
• CLIPS
• expert system
• Gemma
• LLM
• LoRA

The effectiveness and ability of Generative Artificial Intelligence architectures is strongly dependent on the quality of the dataset on which they were trained. Yet, in specialized or technical domains, collecting a large and carefully curated dataset continues to be an obstacle, as publicly available data is often limited, expert annotation is an expensive process, and it might lack verifiability or readability.
The core idea of this project is to leverage the symbolic reasoning and explicit rule-based structure of an expert system to generate semantically coherent and contextually relevant examples, thus representing the critical knowledge of the target domain while avoiding the need for external data collection, and finetune a Gemma3-4B-it LLM.
Setting up a scenario about HACCP food preparation safety decision tree, we shape it like a classification problem to compare the accuracy of our improved model against the untrained version with some pre-defined cases to a scenario described by natural language.
The modeling approach frames a new direction for hybrid AI architectures that are both grounded in human-understandable knowledge and statistical-learned patterns, and aims to be part of a broader objective of creating effective and explainable intelligent systems.