Pillarstone

Nine plants, nine production systems, one model

Pillarstone casts precision components at nine plants in the UK, Poland and Norway, each running its own fork of the same manufacturing execution system. We consolidated them onto one production model without forcing nine shop floors to work the same way.

Client: Pillarstone
Sector: Manufacturing
Run: 11 months · Nov 2023 – Sep 2024
Filed: 2024-12-09
Capability: Systems integration · Data engineering · Cloud migration
Stack: Python 3.12 · FastAPI · PostgreSQL 16 · Apache Kafka · MQTT · dbt · Kubernetes · AWS · Grafana

Impact

9 → 1
Plant-local MES forks replaced by one production model
3 hours
Group production close, down from six working days
4.1%
Scrap rate across all nine plants, down from 6.8%

Challenge

Manufacturing
11 months · Nov 2023 – Sep 2024

The nine forks had drifted for over a decade, and so had their definitions: a "unit produced" was counted at a different point on the line at every plant, which made group reporting an argument rather than a number. Closing production across the group took six working days of manual Excel consolidation, so the board reviewed a picture of the previous month in the middle of the current one. Two plants had upgraded their MES version and could no longer exchange files with the other seven. Any solution that required every plant to adopt one process would have stalled — the Norwegian plant’s traceability requirements are genuinely different from the Polish plant’s.

Solution

We separated the model from the process. One canonical production model defines what an order, a batch, a scrap event and a machine hour mean at group level; per-plant adapters translate local practice into it, so plants kept their own working methods and stopped arguing about definitions. MQTT collectors on 340 machines publish counts directly rather than relying on end-of-shift keying, buffering seventy-two hours locally so a network drop never loses a shift. dbt models the semantic layer, so a metric is defined once and every report derives from that definition. We rolled out plant by plant, smallest first, and left each plant with a named local owner trained on the adapter before moving on.