𝗥𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻 & 𝗗𝗿𝗶𝗳𝘁 𝗗𝗲𝘁𝗲𝗰𝘁𝗶𝗼𝗻 𝗳𝗼𝗿 𝗟𝗮𝗿𝗴𝗲 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀

As LLMs become increasingly integrated into real-world applications, we face urgent questions: • Can we trust their output over time? • How do we measure hallucination without ground truth? • Can we detect silent degradation across model versions like GPT-4, Claude, Mixtral? This project proposes an end-to-end research and evaluation framework that simulates: • 💬 Prompt-level hallucination scoring • 🔄 Drift visualization across model versions • ⚙️ Latency vs. accuracy tradeoffs for real-world inference 🧪 Fully implemented in Python & Streamlit with realistic dummy data 📊 Modular pipeline + figures + scoring functions 📁 GitHub repository: 👉 https://github.com/epaunova/phd-thesis-evaluation-llm 🌍 𝗪𝗵𝘆 𝘁𝗵𝗶𝘀 𝗺𝗮𝘁𝘁𝗲𝗿𝘀: As a Data Scientist and AI Researcher, I believe we need reliable, transparent, and production-oriented evaluation methods – especially as GenAI systems scale across industries. This project is part of my long-term academic and applied research path in LLM safety, alignment, and interpretability. LLM, Drift Detection, Model Evaluation, AI Alignment, Data Science, PhD Thesis