ML Engineer - video surveillance and real-time video analytics
CEREBRO INNOVATION TECHNOLOGIES
Sep 2024 — PresentBuilt and deployed five computer vision models for real-time video analytics, including aggression detection, attribute recognition, tracking, people counting and face detection; developing an LLM assistant MVP.
• Built and shipped to production an aggression classifier on 3D-CNN (X3D-S, 13-frame RGB clip), trained on ~3000 domain videos. Aggression-class metrics: accuracy ~80%, precision 0.83 / recall 0.79 / F1 0.81; runs at 15-25 FPS as a standalone inference service. • Built a more accurate variant (MVP): heavy X3D + lightweight pose model as an event trigger - pose runs continuously (cheap) and launches X3D only when people get close, avoiding running the heavy model 24/7 on thousands of cameras.