Engineering notes
Methods, benchmarks and things that did not work.
Written by the engineers who ran the projects. We publish failure modes as readily as results, because negative results are the most undersupplied thing in industrial AI.
Virtual commissioning: where the schedule savings actually come from
Where virtual commissioning schedule savings actually come from — the three logic failure modes it catches, and the on-site work it does not touch.
Why most machine vision projects fail at the lighting stage
Deep learning cannot recover contrast that was never captured. A practical guide to matching illumination geometry to defect physics.
Collaborative or fenced: a decision framework that is not about the robot
The collaborative vs fenced decision is set by risk assessment, floor space and cycle time — not by the robot's marketing category.
Your industrial AI project will be decided by the data contract, not the model
Most manufacturing AI projects fail before modelling begins. A readiness checklist to run before you commit any budget.
Tolerance stacks in fixture design: the calculation that prevents the argument
Fixture repeatability problems usually trace to a datum scheme that was never analysed. How worst-case and RSS stacks change the design.
Physical AI: an honest boundary for when learned policies beat programming
Learned robot policies solve a real but narrow class of problems. A practical test for whether your application is on the right side.
Topics
Categories we write in.
Next step
Want these in your inbox?
We publish roughly monthly — engineering notes only, no product announcements. Subscribe in the footer, or write to us directly.