Infrastructure Stack

Five layers forming a complete AI infrastructure foundation

5

Agent

AI agents that reason, plan, and execute tasks using the infrastructure layers below.

4

Framework

Model frameworks, serving systems, and distributed training for AI workloads.

3

Compiler

MLIR-based compilers and code generation for optimal model execution.

2

Virt

Virtualization stack from hardware (KVM) through hypervisor (VMM) to containers (Sandbox) for agent isolation.

1

Silicon

GPU architecture and design principles for understanding hardware acceleration of AI.

Why this stack? Each layer exposes the next bottleneck for AI workloads. By building from silicon up, we understand hardware constraints (memory bandwidth, compute), virtualization overheads (isolation, context switching), compiler optimizations (operator fusion, memory tiling), framework efficiencies (kernel selection, communication patterns), and ultimately agent capabilities (reasoning, tool use).

Project Tracker

Status and progress across all layers

CategoryProjectDescription
SiliconNano GPUSIMT + DSA architecture providing Scalar, Vector, Cube / Memory ISA APIs with Linux kernel drivers.
Nano KernelGenKernel generation framework for optimizing tensor computations.
VirtVibe KVMEducational implementation of the Linux KVM API for learning Intel VT-x virtualization at the hardware level.
Vibe VMMMinimal Virtual Machine Monitor (VMM) in C with clean hypervisor abstraction for cross-platform VM execution.
Nano SandboxEducational OCI-compatible container runtime in C with pure-container and Firecracker-backed execution modes.
Nano K8S TutorialHands-on tutorial for building a minimal Kubernetes cluster from scratch using containerd and the K8s control plane APIs.
FrameworkNano TrainLearning-first distributed LLM training framework built around Megatron-style parallelism, covering TP, PP, EP, DP, ZeRO-1/2, and mixed precision on a DeepSeek-style model stack.
Nano TorchDeep learning framework with PyTorch-like API.
Nano ServingModel serving and inference optimization system.
Nano Agentic RLAn agentic RL post-training framework.
AgentNano CoderMinimal terminal-based coding agent with tool use, streaming answers, MCP support, and local subagents for code execution and development workflows.
Nano AgentEvolveA Self-Evolving Agentic Framework.