Model Quantization for Production-Level Neural Network Inference