[+] MANUSCRIPT IN PREPARATION: Unified Framework for Evaluating Vision Models for Action Recognition in Surveillance Context.

[+] Engineered a scalable synthetic-data engine using open-source datasets (SMPL-X, AMASS, and BABEL) and implemented an auto-labeling mechanism (pose/bbox/mask), which accelerates CV/VLM data collection.

[+] Developed a DORI-aligned surveillance dataset (200k+ images & 630k+ QA pairs) with controlled context parameters to enable robust benchmarking of presence & action recognition models.

[+] Built a unified evaluation framework for human-presence & action recognition, with Hydra configs and Docker/CUDA to benchmark YOLO models against compact LVLMs.