# About
I turn machine-learning theory into working, reproducible systems. My research asks a sharp question about causal video generation and video world models: when does spending expensive per-frame compute only on "hard" temporal segments actually pay off? I build open, auditable artifacts — code, locked protocols, and every result file — so each claim can be checked.
Two first-author preprints (one under review at IEEE TCSVT) and the JevGAG + DDAG programme: a matched T4-vs-A100 study of adaptive temporal granularity, an intrinsic pre-generation probe on the generator's own activations, and a training-free seam repair.
# Publications
Regime-Aware Temporal Granularity for Causal Video World Models
When does adaptivity pay? A matched T4/A100 audit of the three preconditions — cost, signal, endpoint — with an intrinsic probe and a regime map.
# Datasets
Open, reproducible artifacts from the JevGAG + DDAG programme.
- JevGAG Audited Results & Figures (E1–E9)
Every headline number re-derives from these files.
- DDAG Adaptive Granularity Experiments
Eleven audited experiments across T4 and A100-80GB.
- SelfAware-DiT Intrinsic Probes
Intrinsic DiT activation probes — AUROC 0.798 vs 0.491 for hand-crafted signals.
- SeamControl Seam Repair
Training-free 25% latent blend cuts the switch-boundary seam by 25%.
- Adaptive Granularity Rollout Videos
Qualitative video evidence on T4 and A100.
- All datasets on the Hub →