Pitstop: AI maintenance agent

Factories often catch the warning signs of a breakdown, but the follow-up is slow and manual: someone reads the manual, checks the parts store, raises a purchase order and finds a free engineer. By then the machine has often failed anyway. A planned tool swap takes 30 minutes; an unplanned breakdown stops a machine for hours. Pitstop closes that gap. A failure model trained on the AI4I 2020 dataset flags a CNC mill whose cutting tool is wearing out and estimates the cutting time left. An agent then plans the whole fix: it finds the repair and the exact part number in the machine manual, checks stock, drafts a purchase order from the cheapest supplier that still delivers in time, and books a qualified engineer after the part arrives and before the predicted failure. If no slot fits, it escalates instead of forcing a bad booking. A manager approves or rejects every draft, and the operator gets a plain-language summary with the downtime saved. The plan adapts when the world changes. Reject a booking with a reason and the agent works around it, keeping the order. If the supplier delays the part, the booking that no longer fits is cancelled and the agent rebooks by itself. In fleet mode, four machines share one spare part and the same engineers without double-booking. The agent runs in our Evolus workspace (Machine Whisperers) on Gemma 4 (31B), served on AMD GPUs through AMD’s Token Factory, with Qwen3.8-27B as an automatic backup. Evolus provides the agent, a knowledge library with the machine manual, and an MCP connector to our factory tools. Guardrails: the agent only drafts, facts come from the tools, timing is enforced, and tool data is treated as data (a supplier quote carrying hidden instructions is ignored). Both models pass all 20 scenarios of our eval, which checks what the agent actually did in the factory systems. The dataset is synthetic; the factory systems, suppliers and people are fictional. Code: MIT, runs with docker compose.