AI Infra

Systems, not recipes. If it is a cluster, a kernel, a serving engine, or a cost/SLO problem, it lives here.

The World Model Platform is the end-to-end case study.

• Performance — rooflines, accelerators, profiling
• Training systems — parallelism, checkpointing, elasticity
• Inference systems — batching, KV cache, disaggregation
• Cluster & platform — Kubernetes, schedulers, GitOps
• Data plane — storage, pipelines, lineage
• Kernels & compilers — Triton, CUDA, LLVM
• Deployment — rollout, serving paths