Rust systems engineer focused on agent orchestration, LLM inference infrastructure, and home-lab automation. Building tools that work reliably at the edge — no cloud dependency, no magic, just well-tested code running where it needs to be.

Agent orchestration framework with execution gating, trust layers, and scope isolation. Pure Rust, designed for reproducible builds and deterministic behavior. The glass abstraction pattern separates stable substrate from disposable integrations.

Portable Execution Substrate for Transformer Inference. GPU-first LLM inference engine using CUDA dispatch with WGMMA/tcgen05 kernels. Currently achieving ~307 tok/s on Qwen2.5-0.5B-Instruct-Q4_K_M (RTX 3070 Ti) vs llama.cpp's 504 tok/s — closing the gap through kernel optimization and weight caching.

Declarative home-lab infrastructure management. Single lab.toml manifest defines machines, services, VMs, and access rules. labctl CLI enforces consistency across the fleet.

Excel replacement for accounting and data workflows. Rust-based PowerQuery-style transformation engine using SQLite instead of XML-zipped binary formats. Consolidates nustage, zed-sheet-lsp, and git-sheets into a unified tabular workflow stack.

• pesti-gguf — GGUF weight loading for pesti
• mirror-log — Event store with decay and attention layers for agent memory
• cargo-declared — Declared vs actual Cargo dependency analysis

• Local-first: Tools should work on your machine, not depend on some API endpoint that might change or die
• Evidence over expertise: Make the code tell you what it does; test limits rather than assume them
• Pragmatic scope: Only enforce limits that are observable or mathematically justified
• Reproducible everything: Same inputs → same outputs, every time

Rust (nightly for SIMD/atomic features), CUDA/C++ interop, Arch Linux, Tailscale networking, dinit init system. Home-lab fleet: x86_64 workstations + ARM Pi5 index node.