Complete archive · 5 notes
Three months of field deployments taught me: the gap between lab and production isn't technical—it's environmental. Temperature swings, vibration, dust, and network flakiness break assumptions faster than any stress test.
Deploy where the work happens. Instrument everything. Build for the conditions that exist, not the ones you wish existed.
Edge compute constraints force clarity. When you have 4GB RAM and intermittent connectivity, every dependency matters. Every model checkpoint needs justification. Every API call becomes a design decision.
Constraints reveal what's essential. The best industrial AI systems aren't the most sophisticated—they're the ones that work reliably under real conditions.
Software can rollback. Physical systems can't. A failed deployment in a factory means halted production. A buggy update to robotic controls means safety incidents.
This changes everything about how you build, test, and deploy. Simulation helps, but it never captures everything. Shadow mode deployments become essential. Rollback procedures need to be mechanical, not digital.
The stakes shape the engineering.
Sensor drift is silent. Calibration degrades gradually. By the time you notice, weeks of data are suspect. Industrial ML doesn't fail loudly—it fails quietly, with predictions that slowly lose accuracy.
Build drift detection into the pipeline. Log calibration metadata. Make sensor health visible. The best models can't compensate for bad instrumentation.
AI safety discourse focuses on existential risk. Industrial AI deals with immediate physical risk. A misclassified defect means shipping bad product. A wrong control signal means equipment damage or worse.
The safety cases are different. Formal verification, redundant systems, human-in-the-loop controls—these aren't philosophical exercises. They're operational requirements. The models need to be right, measurably, every time.