Deployment Engineering as a Discipline
September 1, 2024
Why getting AI systems into production environments deserves its own field
#industrial-ai
6 entries
Letters
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.
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.
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.