PulseGuard

The problem. Unplanned industrial downtime costs up to $125,000 per hour. Plants have sensors everywhere, but alerts without explanation don't get acted on — operators get noise, not root causes. What PulseGuard does. It watches multivariate compressor telemetry (vibration, current, temperature, oil pressure, airflow) and closes the full loop: 1. Detect with math. Rolling z-score over a full daily cycle (so seasonal drift never false-fires), consecutive-hit event grouping, and cross-channel correlation that turns five alarms into one root-cause cluster. Pure numpy — every number is auditable, satisfying the "explain how the prediction was calculated" requirement by construction. 2. Explain with AMD. An open model served on AMD Developer Cloud (ROCm, vLLM, OpenAI-compatible) narrates only computed facts plus cited sections of the maintenance manual. Output is strict JSON: headline, confidence, citations, recommended action. The model can never invent an anomaly — and a deterministic fallback keeps the pipeline alive if the endpoint blinks. 3. Dispatch with Evolus. One click drafts a work order pre-filled with asset, severity, evidence chain, and action; a human approves it; the Evolus workflow executes. High-risk decisions stay accountable.