Associate Product Manager at Delivery Hero, Berlin — logistics CX products at scale.

Off the clock, I don't just use AI agents — I'm building the factory that runs them. I've been designing the pieces of a local-first agent stack: a coding agent, its memory layer, an orchestrator that runs on my own GPU, and a type-checker that turns an agent's authority, cost, and lineage into compile-time facts. I direct the agents to build it while I own the architecture, the guardrails, and the trade-off calls. I can't write the code by hand — but I can specify it, direct it, and tell you exactly why every trade-off went the way it did.

A private, local-first agent stack I've been designing and directing — happy to walk through it in a conversation:

• The agent — an LLM-agnostic, approval-gated coding agent in Rust: multi-provider routing with logged decisions, a fallback tool-protocol for models without native tool use, git-shadow-ref checkpoints and one-command undo.
• The memory — a git-backed, local-first knowledge base and shared-memory contract that agents read and write over MCP.
• The guardrail — an experimental agentic language whose type-checker won't let an agent compile if it exceeds the authority or budget it was granted. Authority, cost, and lineage become compile-time facts, not runtime hopes.
• The orchestrator — a worktree-per-agent design that assigns work to a bench of specialists, with cost accounting and provenance built in.

Public repos where I've built out pieces of the idea:

llm-autobench — Autonomous benchmarking harness: a 7-step discover → pull → bench → judge → report → delete → commit cycle on a cron, using a free NVIDIA NIM judge to keep my RTX 5070 free for the model under test. The commit history is the pipeline running itself.

tether — An installable iPhone PWA that talks to my local + cloud models over Tailscale. No backend, no App Store, no Mac — editor, GitHub commit, and multi-chat, all browser-direct.

changelog-genie — Commits and PRs → clean, categorized changelogs via local LLMs. Packaged as a GitHub Action, 51 tests; nothing leaves your machine.

benchviz — Turns messy LLM benchmark JSON into a static dashboard: leaderboard, capability radar, model×task heatmap, and regression detection that flags any drop >5%. One dependency, pytest-green on Python 3.9 / 3.11 / 3.12, zero API keys.

neon-protocol-ide — Desktop IDE that opens a codebase as an interactive architecture map with conversational AI. 21 releases; IPC as the security boundary so API keys never touch the renderer.

skillich — 1,028 skills across 88 roles, each rated for AI impact. Python SDK + MCP server + OpenAI/Anthropic adapters + CLI. Agent-first by design.

Also public: pulse (Rust/Tauri GPU monitor), rag-starter, mcp-studio, sprint-narrator, decisionlog.

• AI-directed development — every project built by directing agents, not by writing the code by hand
• Local-first — everything runs without cloud dependencies or API keys
• Structural over advisory — guardrails enforced by the type system, not by docs you hope people read
• Documented trade-offs — a DECISIONS.md in every repo: what was chosen, what was rejected, and why