Glassmind

Glassmind is a full-stack explainable AI (XAI) chat application built as a part of my Bachelor's thesis at RPTU (ex TU Kaiserslautern). It runs Meta's Llama-3.2-1B-Instruct and exposes not just a chat interface but a transparency layer into the model's reasoning: for any response, users can inspect Gradient×Input attribution (via conservative propagation through the transformer's norm and attention layers), multi-head attention weights across early/middle/late layers, and Leave-One-Out word importance (computed by re-running generation with each prompt word masked and measuring the drop in output probability).

The backend is a single FastAPI service that loads the model once at startup and serves both the API and a dependency-free React frontend.