ForgeGuard AI – Maintenance Triage on AMD

Factories have the data but not the time: a technician facing thousands of sensor readings and scattered maintenance notes must decide which machine to inspect first. ForgeGuard AI is a decision-support assistant for maintenance technicians and supervisors. How it works: sensor readings and notes (CSV) are validated, unusual readings are detected with rules and statistics (limits plus a z-score against each machine's own past), matching notes are linked, and machines are ranked HIGH / MEDIUM / LOW / OK. For each flagged machine, AI on AMD Token Factory (DeepSeek-V4-Flash, fallback MiMo-V2.6-Flash) writes findings, possible causes, recommended checks and uncertainty. Every AI statement must cite evidence IDs that exist in the case; a verification step checks this, retries once on a bad reply and marks anything unsupported. If the AI is unavailable, a rule-based summary is shown. Measured results: 6/6 planted problems ranked correctly on a synthetic fleet; 4/4 real air-leak failures detected on the MetroPT-3 metro compressor dataset (one held out); 0.76 false alert episodes per week; 4 real AMD responses with 31/31 statements grounded in evidence, about 30 s each. The live web app has a ranked queue with an impact summary, search and voice search, clickable evidence, and attachments for technician photos, video, audio and documents. ForgeGuard suggests what to inspect - a qualified technician decides.