Anurag Bheemani

Defined computer-vision use cases with Vantara stakeholders, built POCs, and iterated on feedback. Trained an attentive probe on a frozen V-JEPA 2.1 model for bear-behavior recognition (75.5% accuracy, 72.6% F1). Deployed on-prem CV inference pipelines on client-controlled GPU/CPU infrastructure to meet strict data-isolation requirements. Reduced V-JEPA 2 inference latency 7.2x by building an FP16 TensorRT engine, then further lowered per-frame latency using NVIDIA DeepStream while reducing GPU memory 39%. Unified Kafka consumer and CV inference into a single container, cutting per-frame latency from 67 ms to 20 ms (70%) and enabling K8s/KEDA autoscaling. Improved leopard detection accuracy from 83% to 92% using DINO-assisted auto-annotation with human validation and shipped inference to NVIDIA Jetson edge devices. Built a semi-automatic annotation platform that reduced labeling time by 70% with model-assisted pre-labeling and lightweight human correction; added multi-user assignment a