The vLLM Moment for Computer Vision: Inside the Architecture of tamserve Aryaman Darda's tamserve decouples visual memory from neural weights in a computer-vision serving architecture, cutting VRAM use by 4x compared with Meta's SAM 2 reference code, which collapsed under multi-camera live feeds. The system is positioned as a vLLM-style serving layer for production computer vision, showing how systems engineering rather than new model weights unlocks live multi-camera tracking. Meta's SAM 2 proved that foundation models can track arbitrary video objects across time, but its reference code collapsed under multi-camera live feeds. Aryaman Darda's tamserve decouples visual memory from neural weights, slashing VRAM by 4x and demonstrating how systems engineering unlocks production computer vision.