With the aim of reducing hallucinations in Large Language Models (LLMs), this project seeks to address this persistent challenge through the emerging field of mechanistic interpretability - a discipline that delves into the inner workings of neural networks to understand why models produce the outputs they do. By dissecting and analyzing internal representations and activation patterns, the project aspires to uncover the mechanisms behind factual inconsistencies and build methods that make LLMs more reliable and grounded. The project has been awarded grants from VCs and ExceptionRaised!