Skip to content

03 Industrial practice

Digital transformation that reaches the shop floor.

Not a portal, not a dashboard programme. Systems that change what a line does, engineered for brownfield reality, segmented networks and people who will judge them on a night shift.

01 Where the value sits

Six places industrial AI reliably pays.

Chosen because the decision is frequent, the ground truth is obtainable, and the improvement is measurable on a P&L rather than in a slide.

01

Vision-based quality inspection

In-line defect detection and classification on surfaces, welds, assemblies and print. Runs at line rate on edge hardware, with a review loop that keeps producing labelled data.

Moves
Scrap and escape rate

02

Predictive maintenance

Asset health from vibration, current, thermal and process telemetry. Ranked by remaining useful life and by what a failure actually costs, not by anomaly score.

Moves
Unplanned downtime

03

Throughput & yield optimisation

Constraint identification across the line, then closed-loop or advisory set-point recommendation. Usually the highest-return work and the least glamorous.

Moves
OEE and first-pass yield

04

Shop-floor data foundation

The unglamorous prerequisite: OPC UA / MQTT collection, contextualisation against asset hierarchy, historian and lakehouse landing, quality monitoring.

Moves
Time to answer a question

05

Digital twin & process simulation

Calibrated models used to evaluate a change before committing the line to it — scheduling, energy, changeover and capacity scenarios.

Moves
Cost of experimentation

06

Agent-assisted operations

Retrieval and agentic assistance over SOPs, maintenance history, drawings and tribal knowledge, delivered where the work happens. Cited, sandboxed, and safe to be wrong out loud.

Moves
Time to competence

02 Reference architecture

Six layers, from the sensor to the decision.

A shape, not a product list. Every engagement adapts it — but the layering, the contracts between layers and the direction of dependency stay fixed.

L5

Applications & decisions

  • Operator UI
  • Planner tooling
  • Alerting
  • Reporting
  • Mobile & HMI

L4

Intelligence

  • Vision models
  • Forecasting
  • Optimisers
  • Twins
  • Retrieval & agents

L3

Data platform

  • Lakehouse
  • Time-series store
  • Feature store
  • Semantic layer
  • Governance

L2

Streaming & contextualisation

  • Kafka / Sparkplug B
  • Stream processing
  • Asset hierarchy
  • Unit & schema contracts

L1

Edge & gateway

  • OPC UA collection
  • Protocol translation
  • Edge inference
  • Store-and-forward
  • Fleet management

L0

Process & OT

  • PLC / SCADA
  • Sensors & vision
  • Historians
  • MES
  • Safety systems

Data and context flow upward · control and intent flow down · contracts at every boundary

03 Design constraints

What industrial software has to survive.

These are not caveats. They are the design inputs that separate systems which get adopted from systems which get bypassed.

Brownfield is the default

Equipment from four decades, three protocols and two acquisitions. We design for what is on the floor, not for a reference architecture diagram.

The line does not stop

Integration work is planned around production windows and validated in a way that does not require faith. Rollback is rehearsed before rollout.

Segmentation and air gaps

Purdue-model boundaries, one-way data diodes and sites with no cloud egress at all. Edge-first designs, offline capable, with sync as an optimisation.

Safety and audit are non-negotiable

Nothing we build sits in a safety-instrumented loop without an explicit, reviewed hazard analysis. Change control records are produced as we go.

Operators decide adoption

A system that gets routed around has failed regardless of its metrics. We design with the shift, not for a stakeholder in a different building.

Skills have to remain in-house

Your maintenance and automation engineers are the long-term owners. Enablement is part of the plan, not a closing workshop.

04 First engagement

Narrow scope, undeniable result.

The shape of a first engagement with a manufacturer. Deliberately narrow — one cell, one constraint, one measurable outcome your own people can verify.

  1. Stage one

    01 Walk, measure, size

    • Process walk with operators and maintenance
    • Data readiness and sensor audit
    • Ground-truth definition and baseline capture
    • Option set with costs and expected effect

  2. Stage two

    02 Prove on your data

    • Working system on real production data
    • Evaluation harness with slice reporting
    • Edge deployment on representative hardware
    • Shadow mode alongside the current process

  3. Stage three

    03 Operate and decide

    • Progressive cut-over with rollback rehearsed
    • Operator enablement and runbooks
    • Measured delta against the baseline
    • Scale-out plan or an honest stop

"An honest stop" is a real outcome. A meaningful share of diagnostics conclude that the economics do not support the project yet — usually because the ground truth is not obtainable at acceptable cost. Establishing that before a programme is committed is the cheapest finding available.

Manufacturing brief

One line, one constraint, one measurable outcome.

Tell us about the line, the data you have, and the number you are trying to move. We will come back with an honest read on tractability and a costed option set — or a recommendation to fix something else first.

Start a technical brief How we engage

or write directly — hello@bluestreaklabs.com