Introduced dual-instrument methodology showing the established leftward LLM bias is instrument-dependent: on abstract questionnaires, 66 models converge center-left; on 48 real Swiss referenda, the agreement gradient flips to center-peaked, with language shifting answers more than political content for some models.
Developed neuro-symbolic system injecting KG concept vectors into LLM embedding space, achieving 272% factual recall improvement over GPT-2 baseline and 213% improvement over RAG with 130x fewer tokens, published at WWW'26.
Built open-source simulator for urban drone traffic management using 3D city models, A* pathfinding, and auction mechanisms to optimize airspace allocation efficiency and fairness.
Implemented Nishita sky model in appleseed renderer (280K+ lines C++) with first-ever ozone layer simulation, achieving superior efficiency over Blender Cycles while maintaining physical accuracy.
.