GearGuard

Mid-sized factories lose hours to machine downtime, yet most stops are never recorded, or are written down vaguely at the end of a shift. Managers cannot see where the biggest losses come from, maintenance stays reactive, and without failure history there is no data to build predictive maintenance on. GearGuard fixes this with a low-cost retrofit. An ESP32-S3 with a clamp-on current sensor, a vibration sensor and a temperature sensor is attached to an existing machine, with no need to cut cables or replace the controller. It streams telemetry over MQTT to a Raspberry Pi edge gateway. When the current drops, GearGuard detects the stop and immediately asks the operator on Telegram why the machine stopped. The operator simply replies with a short text or voice note. The language and document AI runs on AMD GPUs through AMD Token Factory. Voice notes are transcribed by Whisper on the edge gateway, then Qwen3.8-27B classifies the answer into a downtime category with a confidence score; low-confidence answers are flagged for a supervisor to review. Meanwhile, an anomaly detector compares vibration, current and temperature against each machine's learned baseline and raises an early warning before a breakdown. Qwen3.8-27B then explains each alert in plain language and suggests actions, citing motor manuals (converted to Markdown with MinerU2.5-Pro) and past repair records. The dashboard shows machine status, availability, a Pareto chart of downtime causes, downtime cost, MTTR and MTBF, and the alert list, including how many seconds of warning came before each stop. Every operator answer is stored as a labeled downtime event, building the failure history a factory needs to train predictive models over time. GearGuard targets factories running older machines without expensive MES or SCADA systems, sold as a monthly subscription per machine. The hosted demo runs simulated machines alongside our live sensor rig, so judges can try the full flow without hardware.