Hamza
Naseem.

I spent 8 years studying classical Aristotelian logic (Mantiq) — a formal system for deductive reasoning refined in the Islamic Golden Age by scholars like Ibn Sina and Al-Farabi. When I first encountered large language models, I noticed something troubling: they can mimic sound reasoning without actually performing it. That gap — between linguistic fluency and logical validity — is what I work on.

Most AI evaluations test whether an LLM can produce a plausible-sounding answer. My research focuses on a more foundational question: can we mathematically verify that an AI's deduction is sound before it is delivered to a user?

With LogicGuard, my co-author Moiz Ali and I adapted Ibn Sina's 11th-century syllogistic logic (Qiyas) into a deterministic premise-validation engine. Evaluated across 525 TruthfulQA benchmarks on open-weight models (LLaMA and Mistral), it caught logical hallucinations with 100% precision and zero false positives.

I am currently extending this logic-first verification model to autonomous multi-agent pipelines and clinical phone operations at Bytelytic and CALLE-AI.