INDUSTRIAL AI
MLOps for Industrial AI: What’s Different on a Shop Floor
Amey Kadle
13 February 2026
6 min read
MLOps on the shop floor has constraints that web-scale MLOps does not: real-time inference, intermittent connectivity, safety implications and an operator who is not your debugger.
The Six Disciplines
- Model registry with hash, lineage, training data snapshot and eval scores.
- Edge deployment with versioned containers and atomic rollback.
- Continuous evaluation on labelled production data.
- Drift detection on input distributions — not just on labels.
- Shadow deployment before any model goes live.
- Roll-back path that any operator can trigger in under 30 seconds.
The Cadence
- Weekly: drift report, eval scores trend.
- Monthly: candidate model evaluation, shadow deployment review.
- Quarterly: production deployment of the best candidate.
- On-demand: emergency rollback if a metric breaches threshold.
Practitioner note
A model in production but un-monitored is a liability. A model in production with disciplined MLOps is an asset that compounds.
Frequently asked
How often should I retrain?
For vision QA, every 4 — 8 weeks when new defect modes appear. For predictive maintenance, quarterly. For LLM-based agents, whenever the eval set regresses.
Amey Kadle
Founder & CEO, Ajinkya Technologies. 20+ years of building MES, ERP and AI systems for India’s most demanding manufacturing plants.