Plant-floor agents. They answer inside your own tools.

For example: After the afternoon shift, the supervisor sees that Press 7 lost output to a 34-minute changeover and slow running after restart.

• OEE, availability, performance, and quality by cell.
• Aggregated OEE for the entire production line.
• Shift-to-date or completed-shift analysis from the historian.
• Current live analysis from MQTT data.
• Good-versus-scrap production summaries.
• Downtime and loss-bucket analysis.
• Cell-to-cell comparisons, such as Press 1 versus Pack 2.
• What most affected OEE, and what evidence would confirm a suspected cause.
• Operator-friendly narratives and handover summaries in Markdown.

For example: Before the weekly stop, the technician opening Mixer 3 sees the seal kit, lockout steps, and the two similar repairs from last quarter.

• A plain-language maintenance question about one asset.
• Maintenance-history records from a CMMS export, historian, or file input.
• Recent asset events alongside the maintenance history when they are provided.
• Recurring work-order, fault, and inspection themes clustered from similar descriptions.
• A timeline with counts by history type and days since the latest work order, fault, and inspection.
• A concise response grounded only in the supplied asset records.
• Explicit data-gap and integration recommendations for missing CMMS, spare-parts, condition-monitoring, or historian data.

For example: When Cartoner 1 stops three times in a week, the maintenance lead sees the pattern around the same photo-eye and schedules its replacement.

• An incident window compared with a baseline window for supplied time-series signals.
• Rolling-mean and delta detection for signals that shifted around the incident.
• Alarm, event, quality, scrap, downtime, and genealogy evidence when those inputs are present.
• A time-ordered evidence table and incident timeline.
• Candidate causes ranked with deterministic source-weighted confidence.
• Markdown findings with evidence identifiers, what would rule each candidate in or out, and stated data gaps.
• Recommendations for the missing namespace topics needed to strengthen an investigation.

For example: During the night shift, the Line 5 operator is alerted to the press hydraulic-pressure alarm while repeated door-open alarms stay out of the way.

• Structured alarm records or raw MQTT, historian, or export-file alarm events within a defined shift window.
• Alarm episodes clustered by source and code.
• Priority, frequency, and recency ranking for the clusters operators should see first.
• Chattering alarms detected when the same alarm repeats ten or more times in an hour.
• Alarm-flood windows detected at twenty or more alarms in ten minutes.
• A concise Markdown triage brief based only on the computed clusters and statistics.
• Coverage and namespace-topic recommendations when alarm fields or sources are missing.

For example: At 06:00, the Line 2 operator receives a handover noting the filler jam, the changeover delay, and the bearing check due after lunch.

• Good and scrap production totals by cell for a defined shift window.
• Downtime minutes grouped by category, plus recorded quality events.
• Downtime reconstructed from MQTT run/stop streams or duration events when records are not structured.
• Historian, broker-capture, or export-file inputs mapped to the production, downtime, quality, and note roles used by the report.
• A coverage check showing the roles and topics found, ignored, or still missing.
• Namespace topic recommendations that close documented data gaps.
• A requested-language handover in Markdown covering what ran, what stopped, and what the next crew should watch.