- Factory-floor model deployment separates where a model is built from where it scores, so each plant runs its predictions next to the line.
- Cisco builds models on H2O AI Managed Cloud and deploys them with H2O eScorer to the factory floor.
- H2O eScorer scores H2O Driverless AI and H2O-3 model pipelines and Python models on standalone instances, outside the platform's own Kubernetes environment.
- Training, versioning and retraining stay in one governed environment while scoring runs at the plant.
The model works in the data center, and the line never sees it
Your data team can build a quality or yield model. Getting that model to score on every line, in every plant, is the part that stalls.
Plant networks are segmented for good reasons. Operational technology sits behind firewalls, some sites run partly offline, and a score that has to travel to the cloud and back arrives after the part has moved on. You are measured on yield, scrap and uptime, and a model that never reaches the line moves none of them.
How factory-floor model deployment works
The pattern is simple to state: build centrally, score locally, update from one place. Model scoring and inference runs at the plant, while training, versioning and retraining stay in one governed environment.
Step 1: move plant data to the build environment
Where a plant network is segmented or air-gapped, data moves through a controlled path such as a jump server, on the schedule your OT security team sets.
Step 2: build the model
H2O Driverless AI or H2O-3 builds the predictive model for quality, yield or failure risk, and documents it.
Step 3: package it for scoring
Each model becomes a scoring pipeline, a self-contained artifact that scores the same way in the plant as it did in testing.
Step 4: deploy to the plant with H2O eScorer
Deploy the model to a remote H2O eScorer instance at the site, from an uploaded file or the model registry, and score in real time or in batch.
Step 5: watch every site
H2O eScorer's model stats show how each deployed model is scoring, so your team sees when a site starts to behave differently and model drift needs attention.
Step 6: retrain once, update every plant
When a line changes, retrain in the build environment and push the new version to each site from the same registry.
An agent reading the production record
An agent can read production logs, energy use and quality results, then recommend line settings for your team to approve, working from the same signals a model would use, but framed as options rather than a single score.
Results
Cisco uses H2O AI Managed Cloud to build models and H2O eScorer to deploy them on the factory floor.