Associate Product Manager at Delivery Hero in Berlin, working on logistics products at scale. Outside of work, I build software by orchestrating AI — designing the architecture, making the trade-off calls, and directing AI agents to write the implementation.

Pulse — GPU performance monitor for systems running both gaming and local AI workloads. Answers "what can my GPU do right now?" with VRAM process maps and headroom indicators. ~10MB installer, ~30MB RAM.

• Tauri 2 over Electron — native Rust backend for GPU polling at a fraction of the footprint
• NVML integration in Rust for direct hardware access without Node.js overhead
• Zustand with ring buffer for time-series state — fixed-size buffer instead of growing arrays

Neon Protocol IDE — Desktop IDE for navigating codebases through visual architecture maps and conversational AI. Electron + Next.js + React, 20 releases shipped.

• Chose IPC as the security boundary — API keys never touch the renderer process
• Multi-provider LLM routing with priority fallback for zero vendor lock-in
• Zustand over Redux — minimal boilerplate, independently testable slices

Skillich — 1,028 skills across 88 roles, each rated for AI impact. Python SDK with MCP server, OpenAI/Anthropic adapters, and CLI.

• Agent-first architecture — primary consumers are AI tools (MCP, function calling), the web app is a generated artifact
• YAML taxonomy over a database so contributors can submit PRs, not SQL

RAG Starter — Your first RAG pipeline. Local with Ollama, no frameworks, no API keys, 15 minutes.

• Framework-free by design — no LangChain or LlamaIndex, so every line is readable
• Custom exceptions over sys.exit() — library code shouldn't kill notebooks or Streamlit
• Groq free tier for Colab instead of OpenAI — beginners shouldn't need a credit card