AI infrastructure / systems notebook
I study how intelligence emerges from systems.
I work on AI infrastructure and follow this question across the stack: how agent collectives coordinate, and how compute, fabrics, memory, and runtime tuning shape what they can do. This is a working index of my notes, experiments, and projects.
GPU clusters, topology, RDMA/RoCE, NVLink, PCIe/CXL, memory pressure, and tuning.
Cloud systems, heterogeneous compute, model execution, workflow placement, and cost-performance.
Coordination, control planes, shared memory, tool use, deterministic workflows, and review.
How models, memory, tools, agents, and infrastructure combine into new system behavior.
01 / writing
Recent notes
- Scaling 1,000 AI Agents on OCI Kubernetes Engine and File Storage
A storage-backed runtime pattern for 1,000 persistent agents on OKE.
- Agent Graph 0.2.0: Parallel Issues, One Deterministic Control Plane
Parallel issue execution with leases, isolation, observation, and bounded control.
- Agent Graph: A Deterministic Control Plane for Multi-Agent Engineering
A control-plane architecture for software that contains agents inside its workflow.
- PCIe over the Network: When a Remote GPU Looks Local
Remote PCIe virtualization that preserves native driver behavior across a network.
- Raising an Agent: From Execution to Self-Evolution
Four capability levels: orchestration, reliable execution, workflows, and method discovery.
Writing archive
- The Contextual Turn: How AI Is Changing Our Understanding of Language
Language, context, memory, and interaction in AI systems.
- Agentify Cloud: When the Agent Becomes the Cloud Runtime
Agent-native cloud runtime design with policy, route intents, contracts, and validation.
- What Is Tau Computing?
Compute, memory, and fabric design for future AI systems.
- Zettascale in Practice: MRC Benefit
Systems-level thinking for AI networking and infrastructure efficiency.
- MRC and the Future of AI Networking
Network architecture for AI-scale communication and resource composition.
- Run Nextflow with heterogeneous computing on OCI
Pipeline-driven heterogeneous computing across Arm, GPU, and x86 cloud resources.
- Cut Nextflow costs by 70% with OCI
Cost-aware workflow placement for heterogeneous pipelines.
- Zettascale in practice: Scaling beyond limits
Infrastructure notes on scaling behavior, benchmarks, and large AI workload limits.
- Zettascale OSU and NCCL benchmark for H100 AI workloads
Benchmarking RDMA and GPU cluster communication for H100 AI workloads.
- The Story of OpenClaw: Learning to Collaborate
Human-guided AI engineering and system building.
- Human in the Loop: Guiding AI to Break the Bus/Network Architecture Barrier
Guided AI exploration for bus, network, and architecture design.
- Human-in-the-Loop Engineering and Vibe Coding
Human direction, agent execution, and engineering review.
- AI Coding BucketFS: A Transactional FUSE Filesystem for Object Storage
A storage and filesystem experiment shaped by AI-assisted implementation.
- AI Networking: How Fractal Scales Beyond Limits
AI networking notes on scaling, limits, and architecture tradeoffs.
- Inside AI Infrastructure, Series II: Benchmarking
Benchmarking as a way to reason about infrastructure behavior under workload pressure.
- A New Architecture Shift: NVIDIA, Enfabrica, and Intel
Accelerator fabrics and the changing shape of AI infrastructure.
02 / projects
Active builds
- Agent Graph
Deterministic multi-agent engineering for VS Code: issue graphs, MCP control, isolated Codex workers, and parallel execution through bounded workflows.
Project pageInstall VSIX - Agent Runner
Server-side multi-agent orchestration with mailbox queues, agenda tasks, groups, resources, and deterministic project state.
Project page - Agentify Cloud
Agent-native cloud runtime with FastAPI, FastMCP, AGENTS.md policy, route intents, JSON contracts, and validation.
Project page - vLLM-MoE GPU Prefetch
Reduces a normal 80GB GPU demand to 40GB while reaching 128.59 tokens/s on an A100 40GB for 26B MoE inference.
Repo
Project archive
- vLLM MoE CPU Offload
GPU-native MoE offload using CPU host memory as the expert-weight bank for constrained GPU memory.
Repo - llama.cpp-MoE
Router-aware GPU expert slots for local MoE inference under constrained GPU memory.
Repo - GPU-native Scheduler for GPU Computing
Patent-submitted scheduling approach for GPU resource management.
- SuperKernel for SuperPod GPU Clusters
Patent-submitted Jupyter kernel architecture for large GPU fabric execution.
Repo - Nextflow IaC Plugin
Infrastructure-as-code orchestration for pipelines across Arm, GPU, and x86 infrastructure.
RepoReference 1Reference 2 - Distributed MCP Protocol for AI-native CDN Architecture
Distributed protocol design for AI-native content delivery and agent coordination.
Demo - PCIe-Net and RDMA over PCIe/CXL
TCP/IP-over-PCIe/CXL and RDMA-style data movement across high-speed interconnect fabrics.
Demo - CXL Switch SoC and Cluster-on-Board
Switch-chip and multi-CPU board-level architectures for scalable AI and HPC systems.
- Multi-rail HPC Computing System
Multi-rail HPC computing architecture for production rendering workloads.
03 / links