Notes from the shop floor, not the press office.
One editorial rule: we publish what we can show. If a number here describes a Jemba deployment, it comes from a blind test on a real line and you can trace it to a case study. If we have not measured something, we write that instead of estimating it.
Recent posts
Loading…
The proof does not live in the blog. It lives here.
Opinion pieces are easy to write. These three pages are the ones we defend in front of an engineering team, and none of them sits behind a form.
273 h of downtime anticipated across six months of blind testing on a four-machine line — with the per-machine breakdown, including the two machines where the model found almost nothing. Read →
977 columns reduced to 10, and 78.1 % material yield inside the identified operating zone against 68.6 % outside it. Read →
Why we calibrate once and replay month by month, what a false alert costs, and why we publish the machines the model missed. Read →
Name one line and one target variable.
That is the whole ask for a first result blind-tested on your own data.
- No
- Detection rates we have not measured on your kind of line.
- No
- Customer names or logos we have not been given permission to publish.
- No
- Competitor scorecards built on figures we cannot source.
- Yes
- Method, trade-offs, and the numbers from the two lines we have actually validated.
