arXiv, 2026
ArXivDoc: an 8,210-document open-domain scientific retrieval benchmark showing text-based representations outperform document-as-image approaches. Contributed data annotation filtering that distilled 4,059 candidates to 547 evidence-grounded queries.
Applies Sketched Isotropic Gaussian Regularization to the actor trunk of a deep goal-conditioned RL network, preventing representation collapse at scale and improving humanoid evaluation success in Brax.
Article · W&B Report · Code
Defense Stock Claim Analyzer. Scrapes tweets about defense stocks, filters out bots, fetches the actual price movement and surrounding news, then labels each claim as exaggerated, accurate, or understated. Tracks account-level credibility to surface who's consistently right and who's full of it.
Fashion recommendation stack on the H&M Personalized dataset: popularity baseline, KNN content-based filtering (product metadata and ResNet50 image embeddings), and neural collaborative filtering. Includes a static demo and FastAPI backend for live inference.
ViT-Small encoder trained with LeJEPA multi-view self-supervised learning and SIGReg regularization on 1280×720 drone footage (DDOS dataset), served via a Flask + Gunicorn REST API with sub-second inference on arbitrary images.
Explores adversarial attacks and defenses around multimodal embeddings (e.g. OpenAI CLIP ViT-L/14). Replace with a precise one-sentence summary when ready.
Steering experiments to reduce sycophancy (linear probes on LLM activations)