I build kernel-leaning systems for AI infrastructure: KV-cache orchestration, memory hierarchy control, low-latency runtimes, and patent-backed hardware–software interfaces.
Most of my work sits close to the machine — Linux control planes, kernel-facing memory and I/O experiments, CPU scheduling and latency behavior, KV-state movement, and research prototypes that make systems ideas concrete.
230+ technical essays — deep dives on Linux internals, memory systems, inference runtimes, and AI infrastructure.
➡️ manishklach.github.io/writings.html
Kernel & Systems Fast Path
linux-kernel-inference-fastpath — Linux kernel and systems fast path for LLM inference: eBPF tracing, runtime hints, cgroups, NUMA/GPU locality, KV-cache memory policies, and TTFT boost.
Memory Fabric
flash-sram-dram-inference-fabric — Predictive SRAM–DRAM–SSD memory fabric for low-cost AI inference, long-context KV cache tiering, and MoE expert staging.
CPU Performance Control Plane
cpuopt-kernel — Safe, reversible Linux CPU performance profiles across CPUFreq, intel_pstate, amd-pstate, cpuidle, thermal, and hwmon backends.
KV-Cache-Aware I/O
kairo-io — AI KV-cache-aware Linux block I/O: decode-priority scheduling, NVMe backend mapping, placement metadata, and kernel tracepoint visibility.
GPU/RDMA Observability
ai-host-observability — Linux-first host observability for GPU and RDMA systems: memory pressure, PCIe, NUMA, IRQ, and host-side failure signals before they become incidents.
KV-CPU Hardware Interface
kv-cpu-driver — Linux control plane, RTL, and FPGA emulation scaffold for semantic KV-cache orchestration. Patent pending (India App No. 202641056309).
Latency Control Plane
kernel-dvfs-agentic-latency — Kernel latency control plane spanning DVFS, cpuidle, IRQs, scheduler, MM, VFS, I/O, and cgroup budgets for agentic AI.
Intent-Aware Attention
intent-attention-kernel — Intent-aware KV execution for agentic long-context inference: semantic block selection, dynamic scoring, KV quantization, and speculative prefetch.
- KV-cache orchestration and memory residency control
- Linux kernel control planes for inference workloads
- CPU, IRQ, scheduler, and latency-path tuning
- Storage and I/O behavior for decode-critical serving paths
- Systems observability for real AI infrastructure
I work on the hard parts of systems for AI — memory placement, I/O paths, scheduler behavior, latency control, and observability. I build across the stack, but naturally gravitate toward Linux, kernel-adjacent interfaces, CPU and memory behavior, and runtime control planes. I pair code with diagrams, RFC-style docs, and architecture-driven writeups to make low-level work legible.
55 repositories across kernel experiments, KV-cache infrastructure, memory systems, network dataplanes, AI runtimes, and performance tooling.
- Writings & essays: manishklach.github.io/writings.html
- Portfolio: manishklach.github.io
- Patent record: manishklach.github.io/patents.html
- GitHub: github.com/manishklach
- X: @OrbitHigher