These are the areas where I spend my thinking time. Each connects to the others — video understanding needs good representations, representations need good objectives, and objectives need good evaluation.
Detection, representation, video understanding
Architectures, training dynamics, generative models
Policy optimisation, reward design
Cross-modal representation and alignment
Sensor streams, anomaly and event detection
Training, serving and system reliability
Latency-bound inference in operational settings
Models under real-world operating constraints
Implementations of published methods and datasets contributed to the community. Each is accompanied by the research context that motivated it.
Paper-derived implementation of VL-JEPA (arXiv:2512.10942v2) — a vision-language architecture that predicts continuous embeddings instead of autoregressive tokens, achieving stronger performance with 50% fewer trainable parameters.
Fully unsupervised video summarization for courtroom recordings. A policy gradient agent selects frames using only raw visual features — no transcripts, no human labels.
Benchmark dataset built from real-world U.S. courtroom trial recordings with synchronous visual, acoustic, and transcript streams alongside legal expert annotations.