Today
Every line produces hundreds of signals: temperatures, pressures, currents, cycle times. Nobody can watch them all, and fixed limits only trip when it is too late. Creeping drift is therefore noticed only when the sample check shows scrap, and by then a whole shift of product is often affected.
An example
On the sauce line at Tarvina, the SC 250 g herb sauce is running. To bring the sauce to 85 °C in the tubular heater, the hot water has had to be a little hotter every day since Monday, and the pressure drop across the heater is rising. All values are still within their limits. The agent flags the drift, shows that it began with a new batch of starch and links a similar case from March: two days later dark particles turned up in the sauce and a whole shift of product was blocked. The shift leader brings the intermediate cleaning forward, and quality has the new starch batch tested.
Your data, wherever it lives
Your process values live today in controllers and process databases, orders, changeovers and batch changes in the MES, samples and test results in the QM system. demi reads, understands and links them on one timeline per line, with your rights. The AI spots drift in minutes instead of at the next sample check, and shift leaders and quality decide what happens.
Where the data lives today
- Machines & sensors
- MES
- QM system
- Databases
How demi solves it
Three layers in this order: connect the data first, then the rights, then the AI.
- 1 · ConnectData Layer
Data sources in this case: machines and sensors, the MES, the QM system and the process database. demi replaces nothing and puts process values, orders, changeovers, batch changes and test results on one timeline per line. Alerts go to the shift; anything is written back only after approval.
- 2 · PermitPermission Layer
Shift leaders and process engineers see their lines, quality sees all test results. Analysis runs by line, order and batch, never by person. A person decides whether to test, block or readjust, and that is recorded.
- 3 · ActIntelligence Layer
The agent learns the normal state of every line from recorded data, per product and tool, and spots drift that is still within the limits. With every alert it shows the signals that stand out, when the drift began and the most similar past cases with their cause. The values come from the measurements; the language model only explains the connection.
How it runs
- 1
demi reads every line's process values together with orders, batches and test results.
- 2
The agent learns the normal state per line, product and tool.
- 3
When a trend starts to creep, it alerts the shift long before a limit trips.
- 4
With the alert it shows when the drift began, what changed at that time and which past cases were similar.
- 5
Shift leaders and quality decide on testing, blocking or correcting.
What changes
Drift caught while the product is still good
Less scrap and rework
Past cases as a pointer to the cause
What does it bring you?
Check it with your own numbers
Every starting value is an assumption you can change. The calculation sits right below, step by step.
Your numbers
Result
+96 hfreed per year
Assumptions and calculation
demi flags drift early and shows similar past cases; you decide on checks and corrections.
4 h → 3 h per case
How we calculate
8 process deviations × 4 h × 25% × 1.0 (manual) × 12 months = 96 h
Value per year, before demi costs
Freed per year+96 h
Assumption: labour cost of €49.50 per hour in German manufacturing in 2025 (Destatis), plus workplace, IT and overheads.
= Value per year, before demi costs
+€6,240
gained per year
96 h ÷ 1,760 h per role = 0.1 full-time roles
An estimate, not a promise. The starting values are assumptions for an industrial company with 150 to 500 people.
Combine with more cases and get the packageQuestions about this use case
How is this different from fixed limits?
Limits only trip when a single value leaves its tolerance. The agent notices when the interplay of several values shifts, although each one is still fine on its own. That leaves time to act before scrap appears.
Does someone have to build a model for each line first?
No. The agent learns the normal state from the data you already record, per line, product and tool. It recognises start-up, changeovers and planned stops from the MES and evaluates them separately. Your process engineers define which signals count.
How does demi keep false alerts down?
Every alert shows how strongly and how long a signal deviates and how similar cases ended before. When the shift confirms or dismisses an alert, that feeds into the next assessment. Your process engineers set the sensitivity for each line.
Does demi change machine settings itself?
No. demi reads the recorded process values and reports. People change settings and block product, at the machine or in the system, with an entry in the log.
