systems-research

Live GPU emulator for High-Bandwidth Flash: applies HBF timing, capacity and thermal effects to a real LLM inference workload while it executes on a real GPU. arXiv:2609.09800

A closed-loop security runtime preventing "The Great Exfiltration" and Indirect Prompt Injection in Autonomous AI Agents.

TileMEM / TilePO: profile-guided tile-level placement/admission for high-throughput MoE serving.

Fine-grained computation offload for off-the-shelf servers in tens of lines — paper (arXiv:2607.02630), code, and every measurement. Overlap accelerator offloads (GPU/HSM/inference) with other requests by rerouting through the server's own suspend/resume machinery, plus a zero-edit LD_PRELOAD fiber runtime.

Transparent fault tolerance for unmodified servers as a byproduct of in-network total-order communication.

A systems research project exploring how Zig changes the design of dynamic runtimes as a high-level Python implementation

High-performance deduplication and LLM cost reduction engine built on Hyperdimensional Computing. 1,151,378 QPS · 91.6% token reduction · 0.25ms · Zero GPU · Rust

Linux network namespace-based transport performance benchmarking framework using tc, netem, iperf3 and optional eBPF instrumentation.

Experimental Linux RFC for an HBF/CXL-era AI memory control plane: runtime hints, prefetch, placement, and tiering.

Experimental Rust research software for explicit authority boundaries, negative knowledge, traceable state transitions, and reproducible falsification POCs.

Research harness for evaluating query-time bounded elimination of reconstructable KV-cache witnesses in long-context transformer inference workloads. Related provisional filing: IN 202641062451.

Experimental Linux kernel patchset and benchmark suite for semantic memory hints in inference workloads. Explores whether user-space intent (streaming vs reuse vs ephemeral memory) can influence reclaim behavior in Multi-Gen LRU (MGLRU).

Production-oriented, memory-bounded cross-model KV-cache transfer with guarded fallback and reproducible research tooling.

Trace-driven research harness for KV-cache hierarchy policy evaluation in long-context LLM inference.

AI KV-cache-aware Linux block I/O: decode-priority scheduling, placement metadata, NVMe backend mapping scaffolds, and kernel tracepoint observability.

Researching tail-latency trade-offs in Java caching: Caffeine on-heap vs Chronicle Map off-heap using reproducible JMH benchmarks.

Deadline-aware KV-cache scheduling for protecting decode-critical request-state under long-context LLM inference pressure.

Single-GPU LLM decode research prototype: paged KV cache, Triton attention, CUDA append, scheduling, shared prefixes, and multi-layer transactions.

Executable software archaeology: Linux 0.01-derived i386 on modern x86-64/KVM, with reproducible evidence and fault injection.

A Kubernetes-style control plane prototype built for concurrency fault analysis. Features an MVCC storage layer (etcd), informer-driven reconcilers, a status scheduler, lease-based leader election, and fault-injection middleware to analyze implementation behavior against race conditions and status update conflicts.