Designed and evaluated world-model architectures for spatial reasoning, improving long-horizon prediction accuracy by ~18–25% on simulated navigation and embodied AI benchmarks.

Led experiments across 10k–100k+ simulated trajectories, developing data pipelines and evaluation metrics that reduced training instability and cut experiment iteration time by ~30%.

Implemented relational spatial reasoning modules inspired by quantum entanglement to model object to object and scene level constraints, reducing combinatorial search overhead and cut experiment iteration time by ~30%; collaborated within a 5–8 person research team on reusable research code and publications.