WRITING
◇ BLOG
Essays and technical notes on research, engineering, and lessons from the lab.
RESEARCH·2026 · 03 · 18·11 MIN
Using Logit Space of VLMs for Attention to Detail
Notes from a DARPA Triage Challenge perception pipeline: why attention maps alone can't tell you what a VLM actually saw, and how projecting intermediate hidden states through the frozen LM head — the logit lens — gets you a classifier that reasons over trauma cues instead of whole images.
▶ CONTINUE READING
RESEARCH·2026 · 03 · 12·13 MIN
Why JEPA Beat ImageNet Pretraining for Spacecraft Pose
Notes from SPACE-HOP: how embedding-predictive pretraining and a Hopf-fibration rotation grid reshaped our 6-DoF pose estimation pipeline for uncooperative spacecraft.
▶ CONTINUE READING