KAITO is a Kubernetes-native platform for deploying and managing large language model (LLM) inference, fine-tuning, and retrieval-augmented generation (RAG) workloads. It implements custom Kubernetes resource definitions (CRDs) and controllers for Workspace, InferenceSet, RAGEngine, ModelMirror, and MultiRoleInference, providing lifecycle management for GPU-backed AI services. The system integrates with vLLM for inference serving, supports LoRA/QLoRA fine-tuning, and includes a RAG engine with vector stores (FAISS, Qdrant) and output guardrails. Architecturally, KAITO is a hybrid Go/Python system. The core operator logic is written in Go using the Kubernetes controller-runtime framework, with CRD validation, reconciliation, and node provisioning (Karpenter, BYO, Azure GPU-provisioner). Python presets handle inference server startup (vLLM, HuggingFace Transformers), RAG service (FastAPI), guardrails, and benchmarking. The system is cloud-provider aware, with SKU handlers for Azure, AWS, and Azure Local (Arc), and supports model streaming from Azure Blob storage. Key entry points include the workspace controller (cmd/workspace/main.go), RAG engine controller (cmd/ragengine/main.go), and preset generators for model configuration. External dependencies include Kubernetes (controller-runtime, client-go), Karpenter for node provisioning, vLLM for inference, HuggingFace Hub for model access, and various vector databases. The project includes extensive CI/CD, e2e tests, and benchmark regression testing. While primarily an application, it also contains data/research elements such as model catalogs, benchmark baselines, and guardrails policy definitions.