AI infrastructure / systems research

Understanding how intelligence emerges from systems.

I am interested in understanding how and why intelligence arises from systems at scale—from GPU clusters and model weights to large-scale agent collaboration. I build these systems and measure how they behave: GPU communication, RDMA networks, memory movement, model execution, and agent-to-agent communication.

Blogs

Expand all blogsShow selected blogs

Projects

  • Agent Graphagent control plane

    Deterministic multi-agent engineering for VS Code: issue graphs, MCP control, isolated Codex workers, and parallel execution through bounded workflows.

    Project pageInstall VSIX
  • Agent Runneragent runtime

    Server-side multi-agent orchestration with mailbox queues, agenda tasks, groups, resources, and deterministic project state.

    Project page
  • Agentify Cloudcloud runtime

    Agent-native cloud runtime with FastAPI, FastMCP, AGENTS.md policy, route intents, JSON contracts, and validation.

    Project page
  • Memory-efficient MoE inferencemodel runtime

    Treats sparse experts as a managed working set: CPU memory holds the expert bank while routed experts move into bounded GPU slots for fused execution.

    Passive offloadActive cache
  • Nextflow Infrastructure as Codecomputational science

    Reproducible DNA and mRNA sequencing workflows that provision stage-specific x86, Arm, or GPU infrastructure and retain task artifacts for inspection.

    ArchitectureCost study
  • PCIe over Networkremote fabric

    Software-defined remote PCIe virtualization where a host can access physical PCIe devices across a network while preserving native driver behavior.

    BlogDemo
Expand all projectsShow selected projects
  • vLLM MoE CPU Offloadmodel runtime

    GPU-native MoE offload using CPU host memory as the expert-weight bank for constrained GPU memory.

    Repo
  • llama.cpp-MoElocal inference

    Router-aware GPU expert slots for local MoE inference under constrained GPU memory.

    Repo
  • GPU-native Scheduler for GPU Computingscheduling

    Patent-submitted scheduling approach for GPU resource management.

  • SuperKernel for SuperPod GPU Clustersgpu fabric

    Patent-submitted Jupyter kernel architecture for large GPU fabric execution.

    Repo
  • Distributed MCP Protocol for AI-native CDN Architecturedistributed agents

    Distributed protocol design for AI-native content delivery and agent coordination.

    Demo
  • PCIe-Net and RDMA over PCIe/CXLfabric networking

    TCP/IP-over-PCIe/CXL and RDMA-style data movement across high-speed interconnect fabrics.

    Demo
  • CXL Switch SoC and Cluster-on-Boardsystem architecture

    Switch-chip and multi-CPU board-level architectures for scalable AI and HPC systems.

  • Multi-rail HPC Computing Systemhpc rendering

    Multi-rail HPC computing architecture for production rendering workloads.

Talks

Expand all talksShow selected talks

Patents

Expand all patentsShow selected patents

Two Oracle inventions submitted in January 2026—GPU resource management and distributed interactive-kernel orchestration—remain unpublished.

Continue the conversation