Tempo-AI

Factories lose money when the wrong machine is fixed first. Tempo helps a maintenance team decide which machine to repair, and on which day, given limited technicians and late spare parts. Forecast. A gradient-boosted model predicts each machine’s remaining life as a calibrated range, not a single guess. On machines it never saw, the median error is 8.2 days, against 17.9 days for guessing the average. Plan. Tempo replays the future thousands of times to price each repair day, then solves the whole schedule exactly with the Hungarian algorithm. On scarce-capacity fleets it is 21.4% cheaper than a tuned alarm. Explain (AI on AMD). DeepSeek-V4-Flash, served through AMD Token Factory on AMD GPUs, writes a plain-language explanation and drafts a work order for each repair, using only numbers from the plan. Check. A deterministic rulebook validates every work order, and every number in the AI’s text is checked against the plan. In a stress test where the model was not told the safety rules, none of the unsafe drafts reached a person. Approve. A person must read the explanation before Approve unlocks. Every decision goes into a hash-chained audit log, so quiet edits are detected. Honest limits: the factory is simulated, the cost figures are assumptions, and the planner runs on CPU today. The AMD usage in this project is the LLM layer. The safety checker matches keywords, which is why a person still approves.