Resources
The questions we actually get asked.
Short answers. Each one points at the page that holds the detail, so nothing here is a second version of the truth.
Getting started
- Q
- Do we need a data scientist?
No. Variable selection, preprocessing, model selection, tuning, calibration and interpretation all run inside the product. How that works → - Q
- What do you need from us to start?
One line, one target variable, one export — plus one process decision: which of your variables are settings and which are results. The two-week blind test → - Q
- We only have a CSV. Is that enough?
Yes. A rough export is a starting point, not a blocker. Both published client cases ran on exports the plant was already producing. What reaches us → - Q
- How long before we see a number?
Two weeks, on a plant-level scope. The long pole in most industrial AI projects is integration, and there is not one here. Why → - Q
- How long does it take?
Two weeks for a first blind-tested result, then four to eight weeks to move to continuous operation on one line. Details →
On the line
- Q
- Do you need to install sensors?
No. No new sensor and no intervention on the machines. The aerospace case ran on the run/stop event log alone. That case → - Q
- Will this disrupt production?
Nothing is added to the line and nothing is installed on your network. Scoring is one lightweight computation per machine every ten minutes. Requirements → - Q
- Does it replace our MES or historian?
No. Jemba reads what those systems already hold and computes the layer neither of them does. Side-by-side → - Q
- What happens after a process change?
Retraining takes one click and about thirty seconds, done by your own team. More →
Trusting the numbers
- Q
- Are your published figures projections?
No. Both are measured — one by backtest on periods the model never saw, one by counting the yield of past runs that fell inside the recommended zone. The method → - Q
- Does it work on every machine?
No, and we publish where it did not. On the aerospace line two of four machines paid off; the other two are in the deliverable and are not priced into the case. Which ones and why → - Q
- What if the answer is that our data carries nothing?
Then you get a documented no, and you keep it. That is a legitimate outcome of the test. How that works →
Data and security
- Q
- Where is our data hosted?
On our own servers in France and the United States, with no third-party cloud subprocessor. Security & data → - Q
- Will our data train models other clients use?
No. Never. Model isolation → - Q
- Are you certified?
ISO 27001 and SOC 2 are in their final stages. We will keep saying exactly that until they are signed. Current status → - Q
- Does this change our GDPR scope?
For anomaly detection, no — micro-stop counts only, no named operator data. Personal data → - Q
- Can we get our data deleted?
Yes — permanently, along with every model derived from it, on request. Deletion →
Not answered here?
Ask us the awkward one.
If your team has a technical or security question we have not covered, send it. We answer in writing.
