LLM evaluation & red-teaming
Systematic benchmarks, adversarial cases, failure analysis and comparisons that make weaknesses visible rather than hiding them behind a single average score.
Research / Evidence & Uncertainty
Research in Trustworthy AI.
I study how intelligent systems can be evaluated rigorously, communicate uncertainty clearly, and remain useful when the evidence is incomplete.
Research directions
Evidence before confidence.
My work connects empirical model evaluation with practical interfaces and engineering workflows. I am especially interested in whether model outputs are supported by evidence, how failure modes are measured, and how systems should behave under uncertainty.
Grounding and hallucination detection
Checking whether generated claims are supported by accessible sources, with citations, traceable reasoning and transparent limitations.
Responsible AI decision support
Designing uncertainty-aware interfaces that support informed decisions without overstating safety, confidence or completeness.
Selected research projects
Hallucination Detection in LLMs
A systematic study comparing eight detection methods using 200 benchmark cases. The best method achieved 90.5% accuracy within that evaluation setup. The work includes a research dashboard for inspecting source evidence and detector results.
Read the benchmark case study →
Autonomous Taxi AI Ethics
An interactive prototype for exploring uncertainty-aware drop-off recommendations, local risks, explanations and safer nearby alternatives. The work focuses on responsible decision support rather than deployed autonomous driving capability.
Explore the ethics case study →
Time-aware recommender systems — in progress
Current exploratory work examines how recommendation methods can account for changing user preferences over time. Temporal evaluation and coverage matter alongside prediction error. Any approach or improvement remains research in progress, not a validated product claim.
How I approach evaluation
Well-designed tests should define the task, evaluation examples, baselines and failure modes before making claims about performance. Results should be interpreted in light of the dataset, measurement choices and resources used. For AI-assisted decision support, model-level tests alone cannot demonstrate improved human decisions; that requires separate study.