Research notes, postmortems, and long-form explanations — mostly ML systems, computer vision, and the occasional hard-won lesson.
Three weeks of controlled RoPE-for-CLIP experiments, one adversarial literature review, and the ICLR 2025 paper that had already published my headline result. Includes a full primer on rotary position embeddings, the SigLIP loss, and NTK-aware length extrapolation.