Pose or pixels? Three ways to recognize fights on RWF-2000
R(2+1)D-18, a Transformer over YOLO11-Pose dynamics, and a two-stream cross-attention model, plus why video-level labels call for motion-aware clip mining.
Writing
Notes on where models fail
Longer write-ups of findings from my research and projects. They're being written now; each one expands on a result described elsewhere on this site.
Setting a citation-check threshold from real failures
When a small LLM writes only from deterministic facts, a simple word-overlap test can catch misattributed citations. The data put wrong citations at about 17% overlap and correct ones near 100%.
A reranker meets financial filings: faithfulness from 0.93 to 0.77
What a general-domain cross-encoder did to a RAG system over SEC 10-K filings, and why retrieval components need domain-specific evaluation too.
0.998 macro-F1 and nothing learned: shortcut learning in fingerprint alteration detection
A perturbation analysis on SOCOFing, and what it shows about how synthetic benchmarks can reward detecting the editing process instead of the task.
Want to know when these go live? Send me an email.