Cognitive Embeddings: Neural Encoding Models as Similarity Metrics
Gershony, D. — Ongoing
- Created a new embedding algorithm using Meta's fMRI brain encoding model to measure how the brain processes media.
- Ran Python, TRIBE v2, and PyTorch experiments across 900+ text pairs; results show brain-based similarity captures fundamentally different signals than standard embeddings (p<0.001); currently training a distilled model for public release.
A Pattern Hierarchy for Using AI for Task Allocation in Agile Project Management
Patel, M., Gershony, D., et al. — Accepted to XP 2026
Advised by Prof. Hadar Ziv and Prof. Alberto Krone-Martins (UC Irvine).
- Built an ML system that auto-routes GitHub issues to the developer most likely to resolve them, trained on 500k+ historical tasks.
- Improved assignment accuracy MRR +74% through CodeBERT embedding-based feature modeling and scikit-learn ranking.
- Accepted to the International Conference on Agile Software Development (XP 2026).