I build AI inference infrastructure and container runtimes, from WGSL compute kernels to distributed GPU clusters, with side quests in binary reverse engineering.

🧠 AI inference, down to the metal

• agave: LLM inference engine written from scratch in Zig, no Python runtime, no framework tax
• beam: Ray's API reimplemented in ~1,400 lines of pure Python, enough to run vLLM multi-node. NVIDIA and AMD, validated cross-node
• vllm-webgpu: vLLM platform plugin targeting WebGPU: WGSL compute kernels instead of CUDA
• muninn-sidecar: Persistent memory for AI agents, shipped as a Go sidecar
• gpustack-modelsync: Declarative model placement across GPU cluster nodes
• gb10-thermal-toolkit: Thermal-driven GPU clock governor for NVIDIA DGX Spark, fixes the under-load power-off
• dgx-spark-deepseek: Field notes from serving DeepSeek-V4-Flash on a 2-node DGX Spark RoCE cluster, everything measured on real hardware
• review-prompts: Auto-fix review loop, 37 specialized prompts dispatched to whatever AI coding agents you have installed

🕹️ Reverse engineering

• rebrew: Compiler-in-the-loop decompilation workbench for binary-matching reversing
• europa1400-networkfix: Fixed multiplayer in a 2001 game the vendor abandoned
• openmiles: Open reimplementation of the Miles Sound System
• resembl: Assembly code similarity search
• recoverage: Coverage dashboard for binary-matching decompilation projects
• europa1400-lua: Lua console injected into a running 2001 game, for reverse engineers and modders

⚙️ Runtimes and clusters

• katamaran: Live migration for Kata Containers
• vmetal-openshift: Virtual baremetal OpenShift lab: Redfish BMC emulation, bonded NICs, split-DNS, one Ansible topology

Zig for things that must be fast, Python for things that must exist by Friday.