A working set of systems where operations and modern AI meet. Each one connects real data to agents that can use it — and is held to the standard of something that has to run, not just demo.
Plant-floor data, retrievable by an agent.
A pipeline that moves operational data from where it actually lives — PLCs and historians — into a place an AI agent can use it. OPC UA feeds a Python service into MongoDB; machine-tag embeddings and plant documents go into a FAISS vector store; LangChain grounds an agent in that data so it can answer real operational questions instead of guessing.
An operating system for an AI-run company.
An open-source operating system for running a company on AI agents — a world model held as a file system, with a human proxy as root. TypeScript and Fastify backend, MongoDB Atlas with Mongoose, BullMQ for job orchestration, Ollama for local inference and Claude in production, deployed on Railway. Built with Claude Code.
A 29 MB model that switches real pins, fully offline.
A proof of concept that splits a small assistant the way a control system splits a job. Cactus Needle, a 29 MB tool-calling model, proposes an action with a confidence score; plain Python checks it against an allowlist and risk-based thresholds and owns the GPIO pins; a 0.5B Qwen model in Ollama writes summaries only when text is needed. Tested on a Raspberry Pi 4 against the same 11 commands in two versions: the best routed 7 of 9 actionable commands correctly, and no wrong physical action ran automatically across 22 test commands. Both scripts are in the article.