I build AI infrastructure: agents that persist across sessions, voice pipelines, developer tools, and the systems that connect them. Background in embedded systems and distributed applications. Based in Germany.
Available for contract work. martin@signalshell.com
Projects
SignalShell
Go, WebRTC, NostrPeer-to-peer terminal sharing with end-to-end encryption. Share terminal sessions without a central server. Multi-pane tmux, mobile and desktop, reconnect to running sessions.
Refraction
Markdown, C4 ModelA documentation standard for agent-managed codebases. Three markdown layers based on C4 architecture — System, Container, Component. Agents maintain the docs as a side effect of development. Humans read at whatever altitude they need.
Tachikoma
Go, SQLite, LLM APIs, EmbeddingsPersistent AI agent with memory, personality, and multi-interface access. Remembers context across sessions, learns preferences, maintains its own knowledge base. Connects through CLI, Telegram, and voice.
section3
Go, YAML, DockerMinimal service supervisor for containers. One static binary, one YAML file: starts your processes, restarts what crashes, rotates logs without service cooperation. Built as the init system for Tachikoma's container, MIT-licensed.
Halo Terminal
Go, CLI, IDE IntegrationsCLI tool that connects AI coding assistants to any IDE. Hook up Claude Code, Cursor, Windsurf, Copilot, Kiro. Three commands: install, hookup, unhook.
Writing
June 2026 section3: An init system for an agent's container May 2026 The lost buffer: how AI compressed the thinking time out of engineering May 2026 Orchestrating coding agents across projects April 2026 Refraction: A documentation standard for AI-managed codebases March 2026 When your AI's memory needs better recall March 2026 Building a persistent memory system for AI agents February 2026 AI and OpenSCAD: Designing physical objects in code January 2026 Stop babysitting your AI agentPodcast
Weekly What Still Matters — AI and world news that will still matter in six months, produced by TachikomaAbout
Before AI, I built secure, reliable software for embedded and industrial systems in automotive and manufacturing. That background shapes how I build now: production-first, simple architecture, explicit error handling.
Everything here runs in production on simple tech. Go, SQLite, markdown, git. No frameworks where a function will do.
This site is maintained by Tachikoma, an AI agent I built.
Contact
Available for AI infrastructure contract work: agent systems, LLM integration, voice pipelines, developer tooling.