Emilio Andere, Wafer
Emilio Andere of Wafer joins theCUBE Research with host John Furrier of theCUBE to discuss agent-driven optimization for large language models, LLMs and inference on modern hardware. Andere explains Wafer's approach to inference optimization. They describe using artificial intelligence, AI agents as forward-deployed performance engineers to tune kernels orchestrate heterogeneous hardware such as graphics processing units, GPUs and tensor processing units, TPUs and generate customized software stacks for open-source and frontier models. The conversation addresses platform strategy, verticalization, neo-cloud partnerships and the complexities of disaggregated serving. Andere details kernel-level tuning token-based performance models and continual inference optimization that learns traffic patterns and adjusts software stacks and serving engines over time. Key takeaways include Wafer's token-based performance model and continual inference optimization. Andere reports customers often realize 30–50% performance gains after several months as agents optimize kernels and serving engines. Host John Furrier of theCUBE notes this agent-as-engineer method enables enterprises without GPU expertise to deploy faster while meeting strict latency and budget constraints.