Case study · Manufacturing
32% less unplanned downtime. 5.8-month payback.
A 12-plant mid-market manufacturer combined asset-class ML with work-order narrative mining — running fully on-prem — to turn decades of CMMS free-text into a predictive maintenance program that planners actually trust.
- Industry
- Mid-market discrete manufacturing
- Region
- US (12 plants, 3 countries)
- Size
- ~2,800 monitored assets
- Stack
- Historian · OPC-UA · Postgres · vLLM (Llama 3) on-prem
- Engagement
- 14 weeks build + 8 weeks per-plant rollout
Results
Measured at 12-month milestone.
32%
Reduction in unplanned downtime on tracked assets
22%
Increase in MTBF across critical equipment
18%
Drop in spare-parts overspend
5.8mo
Time-to-payback on the program
The challenge
Data was siloed, narratives were unread, trust was burned.
- 01
Five different historians across 12 plants, none of which the maintenance team could query easily. Reliability engineers spent most of their time pulling and reconciling data instead of acting on signals.
- 02
Decades of work-order narrative — what technicians actually did during service — were locked in CMMS free-text fields. The most valuable diagnostic context for predictive models was the part nobody was reading.
- 03
A previous platform pilot had failed because the model recommendations didn't match what experienced maintenance planners knew about the equipment. Trust was burned.
- 04
Strict OT/IT separation rules: nothing on the production network could phone home to a cloud API. Edge or on-prem only.
The solution
Four capabilities of the predictive maintenance agent.
01
Unified historian + work-order corpus
ETL from 5 historians into a single time-series store. CMMS work-order narrative ingested with the structured fields (asset, codes, dates), indexed for retrieval, and tagged with asset-class taxonomy. Single queryable surface for sensor + narrative.
02
Asset-class predictive models with narrative grounding
Class-specific anomaly models (motors, gearboxes, hydraulics, conveyors) producing structured probabilities. An agent layer combines those probabilities with retrieval over recent work-order narratives on similar assets, then drafts a maintenance recommendation with explicit reasoning a planner can evaluate.
03
Planner workflow with explicit trust ladder
Recommendations land in the planner's queue, not auto-actioned. Each one shows the underlying signals, the comparable historical cases, the predicted failure window, and a confidence band. Planners approve, defer, or reject — every decision logged and used for next-iteration tuning.
04
On-prem inference with edge agents
Llama 3 70B running on plant-side GPU cluster (vLLM). Lightweight edge agents per plant ingesting historian and CMMS data, sending anonymized signal patterns (not raw data) cross-plant for benchmarking. No data leaves the OT boundary.
"Our last vendor pilot died because the recommendations were generic. This one earned trust on the floor in 90 days because planners could see the reasoning — and the reasoning matched what our senior techs already knew."
VP Reliability Engineering
Mid-Market Discrete Manufacturer (12 Plants)
Related services
AI Agent Development →Production agents for asset-heavy industries, including on-prem deployment and OT/IT-respecting architectures. From $80,000.