Diversity-Aware Retrieval for RAG
Dual residual tracking treats retrieval as coverage, not ranking: nDCG stays competitive (0.561 vs 0.612) while redundancy drops from 0.770 to 0.406.
Interested in how things work, and how they could work better.
President @ CS + Social Good
Math, CS & Public Policy @ Stanford
Turning data into models, software, and decisions. Understanding how real-world systems work and how technology shapes them.
Dual residual tracking treats retrieval as coverage, not ranking: nDCG stays competitive (0.561 vs 0.612) while redundancy drops from 0.770 to 0.406.
Completeness of compact subsets; compactness and isometry under Gromov-Hausdorff distance.
CNN + temporal heuristics for real-time guitar strum detection from gestures.
Properties from differential equations; applications in quantum mechanics.
Shamir vs Blakley; threshold schemes for secure distribution.
Directed and edited opening ceremony for 1,500+ participants.
Data viz + ML for environmental sensors; health factors in the Philippines.
When reordering terms preserves convergence or yields any limit.
ML on news transcripts; political reporting patterns via TF-IDF and clustering.
BERT-based analysis of Reddit; topic correlations across partisan discussions.