I build the machinery around models.
I design and prototype AI systems from the Linux and storage layers up through GPU kernels and inference control planes. My work focuses on a practical question: how do we make large-model inference faster, cheaper, and easier to operate?
The repositories here are executable architecture work—code, benchmarks, simulators, kernel RFCs, hardware interfaces, and production-oriented reference systems.
| Project | What it explores |
|---|---|
| Inference Factory Simulator | Trace-driven LLM serving economics: continuous batching, disaggregated prefill/decode, speculative decoding, MoE balance, tiered KV reuse, GPU cost, and margin. |
| ExpertMesh | A low-cost AMD/ROCm stationary-expert fabric for trillion-parameter sparse MoE inference, with packed-INT4 HIP kernels and activation-only RoCE transport. |
| KVSSD Attention | SSD-resident INT2/INT4 KV cache with bounded asynchronous staging and fused CUDA/Triton dequantizing decode attention. |
| K3 Inference Platform | A production-oriented MoE serving control plane: checkpoint release gates, capacity planning, admission, OpenAI-compatible routing, benchmarks, and observability. |
| GB300 RL Runtime | A close-to-metal C/CUDA runtime with persistent GPU workers, hugepage KV arenas, command rings, and no per-token CPU orchestration. |
| AttnRes Kernel Lab | Fused PyTorch/Triton kernels for routing across model depth and sequence time, including Hydra-2P, KDA, and CADR. |
| Linux HBF Control Plane | An executable Linux memory-tiering RFC with asynchronous range hints, NUMA migration, tracepoints, selftests, and observability. |
| OpenFlash Controller Lab | An executable NAND-controller architecture spanning FTL, scheduling, queue ABI, QEMU PCI emulation, and a Linux blk-mq driver. |
model algorithms fused attention · routing · low precision
serving systems admission · scheduling · release gates · economics
GPU runtimes persistent execution · command rings · fast paths
memory systems HBM/DRAM/NVMe tiering · KV placement · CXL/HBF
host + kernel Linux MM · block I/O · NUMA · eBPF · observability
hardware interfaces flash controllers · RTL scaffolds · CPU/GPU fabrics
- Treating inference as a factory, not a single kernel: utilization, placement, queueing, cost, and operational evidence all matter.
- Moving fewer bytes: stationary experts, compact activations, quantized KV state, semantic reuse, and predictive residency.
- Removing work from the token path: persistent GPU execution, bounded queues, preallocation, and explicit control-plane boundaries.
- Making architecture claims testable through benchmarks, simulators, CI gates, tracepoints, and honest maturity labels.
- Linux inference fast path — eBPF, cgroups, NUMA/GPU locality, KV memory policy, and TTFT control.
- AI host observability — Prometheus signals for GPU/RDMA hosts, PCIe, NUMA, memory pressure, IRQs, and kernel events.
- RL inference scheduler — DQN scheduling against FIFO, SJF, and priority baselines.
- CPUOpt Kernel — safe, reversible CPU performance profiles across Linux power and thermal backends.
- MLX Metal Kernels — experimental Apple Silicon kernels for attention, decode, and KV-cache primitives.
I pair systems code with architecture diagrams, RFC-style documents, and long-form technical writing. The portfolio includes 230+ essays and a record of 68 patents filed or granted worldwide across memory, storage, runtimes, and hardware–software interfaces.
Read the essays · Browse the portfolio · View the patent record · Follow on X