01
Mold
Housing, 12 s
China line · lot L-1045 · 2,400 pin headers
01
Housing, 12 s
02
Press P2
03
Crimp
04
Camera
Lot L-1045 is about to enter mold.
Why agents
A traditional line moves the lot. An agent moves the decision, and only when the plant can carry it out.
The press current is already in the historian. The camera raises six alarms. The operator clears them, because the lot is late, and L-1045 leaves the line.
Containment starts after the shipment. Four hours to name the lot. Press P2 is down 38 minutes. $432 of this lot is already gone.
The same drift reaches a repair check: an anode basket, and a plater on this shift. Stock is 0, so no work order opens. The press keeps running.
The same six alarms become one incident: plating void, pin 3, Inspect. The hold waits for a second person. The lot stays. Four minutes to contain. $0 of this lot ships.
The apps on this lot
Computer vision, predictive maintenance, inventory, planning, and OEE each report one fact. The agent is the role that uses those facts on L-1045, in one decision.
The camera sees a plating void on pin 3. The agent turns six alarms into one incident at Inspect.
The drift is the cause. The agent asks for a repair only when a plater is on this shift.
Stock is 0. The agent leaves the work order closed, and the press keeps running.
The schedule wants L-1045 out the door. The agent holds it at Inspect, so the defect is settled first.
The void is contained in 4 minutes. Press P2 downtime stays at 0, because the repair cannot be done.
The agent drafts the hold. A second person approves it. Release uses the same two-person rule.
Click a system to inspect it — then ask "what if it's missing?"
Telemetry and events exactly as captured in Databricks — immutable source-of-truth layer.
Cleaned and validated by Foundry transforms — conformed schemas, ready for joins.
Enriched and aggregated — powers ontology objects, features and every application.
No generations archived yet — a new generation is captured every 6 simulated hours.
sensor_7.raw → databricks bronze → foundry transform → gold → ontology object · 0.8 s| Table | Source | Rows | Format |
|---|---|---|---|
machine_telemetry | Edge / historian | 0 | Parquet |
alarm_events | SCADA | 0 | Parquet |
order_progress | MES | 0 | Parquet |
quality_events | MES | 0 | Parquet |
shift_reports | MOM | 0 | Parquet |
material_master | SAP | 0 | Parquet |
ontology_objects | Foundry → Ontology | 0 | Objects |
ml_predictions | ML · Mosaic AI | 0 | Parquet |
| Order | Product | Qty | Progress | Status | Due |
|---|
| Area | Without digitalization | With digitalization | Gap closed |
|---|
| ID | Machine | Type | Priority | Status | Description | Parts |
|---|
Select an object to inspect its properties, links and action history - then run an action.
Select a machine to inspect it, or run a what-if scenario.
| Machine | Line | P24h | P48h | Risk | Recommended | Model card |
|---|
● critical impact · ◐ partial impact · ○ no direct impact — when the component on the left is missing
| Date & time | Actor | Action | Detail | Outcome |
|---|
Machines, PLCs and meters stream telemetry to edge gateways.
SCADA, MES, MOM and SAP exchange data automatically.
One lakehouse for storage and compute; Foundry unifies it into governed datasets.
Live objects with actions power apps — and send decisions back to operations.
| KPI | Manual | Digitalized |
|---|---|---|
| OEE | 68% | 91% |
| Throughput | 1.0× | +18% |
| Reporting latency | 2 days | <1 min |
| Scrap rate | 3.4% | 1.1% |
| Batch traceability | 2 days | seconds |
| Unplanned downtime | 7.1% | 4.8% |
Databricks stores the lakehouse; Foundry unifies it into governed datasets with lineage — no more silos and spreadsheets.
Machines, orders and batches become live linked objects — simulate changes on the twin before touching the line.
Batch → order → machine → material reconstructed in seconds via object links, satisfying audits and recalls.
ML trains on Databricks Mosaic AI; Foundry agents turn predictions into work orders and dispatch changes automatically.
Role-based dashboards and alerts fire in real time — planners act on live truth instead of weekly meetings.
Automated integration removes paper logs, Excel re-entry and phone calls — and the errors they create.