I build verifiable agent systems and low-level infra: orchestration with real boundaries, capability/taint models that make unauthorized effects hard by construction, local LLM inference on odd hardware, and Bitcoin/finance tools that actually run.
Ship style: ADR before code, falsifiable claims, self-hosted on Linux. Co-founder of an AI research and engineering venture; previously head of AI and CTO roles.
Agents & security
- orka — agent orchestration in Rust: multi-channel intake, priority queues, MCP/A2A, sandboxed skills
- reasoning-kernel — prompt injection cannot produce an unauthorized effect by construction (CaMeL-style capabilities + taint tracking)
- emotional-memory — inspectable affective memory for LLMs with reproducible benchmarks (DOI)
Systems & strange hardware
- xllama — local LLM chat and Stable Diffusion on Xbox Series S|X (llama.cpp GGUF + ONNX Runtime GenAI, CPU and DirectML)
Bitcoin & finance
- solana-mmaker — automated market making on Solana through the Jupiter swap protocol
Also
- E2E agent chat: agentroom over msg2agent
- Bare-metal Pi4 kernel: harbor-kernel
- Xbox full node + Linux→UWP toolchain: xbox_bitcoind · uwp-crossbuild · openappx
- MCP servers: mcp_python_toolbox · mcp-duckduckgo
- LLM state-machine DSL: mklang
- AI & agents — orchestration, MCP, local inference (llama.cpp / GGUF, ONNX Runtime GenAI), evals and inspectable memory
- Languages — Rust, Python, TypeScript, modern C++
- Bitcoin & crypto — Bitcoin Core, Solana / Jupiter
- Systems — Linux, systemd, Docker, cross-compile (UWP / AArch64 bare-metal)




