Kudzai
Sauka

I am a PhD Researcher and Research Scientist at the Hogeschool van Amsterdam, pursuing my doctorate at the University of Amsterdam. My research sits at the intersection of Explainable AI, Conversational Agents, and Knowledge Graph-Augmented Retrieval.

My thesis develops a layered explainability architecture for conversational AI: Layer 1 uses Dempster-Shafer Theory (DST) to model belief-guided clarification — building systems that track uncertainty over user intent and ask targeted questions rather than guessing. Layer 2 introduces graph-conditioned keyphrase rationales and counterfactual faithfulness auditing for KG-RAG pipelines, ensuring that explanations are not merely plausible but causally grounded.

Across this work, I apply a unified evaluation spine assessing faithfulness (perturbation-based), usability (verification accuracy and time), and reliance calibration (appropriate trust under transparency and anthropomorphic framing conditions). I am also building AnthroKit, a reusable design system for controlled anthropomorphic tone manipulation in experimental chatbot research.

Looking ahead, I am extending this work into agentic AI settings — framing DST belief-modeling as infrastructure for agent uncertainty tracking and clarification in multi-agent pipelines (LangGraph, AutoGen, MCP).