Inspection is still sampling
One in fifty checked. Hope for the rest.
Home / Industries / Manufacturing
Manufacturing
Inference on the line, not in someone else's datacentre.
Vision inspection, predictive maintenance and plant analytics that run at the edge.
- Plant-floor edge compute
- Every unit inspected
- Survives network loss
The starting point
What we hear in the first meeting.
If none of these sound like you, we are probably not the right call yet — and we will say so.
Downtime is discovered, not predicted
The schedule is a calendar, not a signal.
The plant network cannot be trusted
A round trip to the cloud stops the line.
Our approach
If it can stop the line, it runs on the line.
Plant-floor AI has a constraint most software does not: the fallback cannot be waiting for the network.
How we work →-
Runs at the edge.
Local inference; cloud only for reporting.
-
Degrades safely.
Uplink drops, inspection continues, syncs later.
-
Built around OT reality.
Your PLCs, cameras and SCADA stay.
-
Measured against today.
We record the baseline before we claim anything.
Proof
How we have solved it in Manufacturing.
Challenge, solution, outcome. Names withheld, numbers not inflated.
Automotive supplier
- Challenge
- Quality checks sampled one unit in fifty and hoped for the rest.
- Solution
- Line-side vision inspection on local compute, with the operator in the loop.
- Outcome
- Inspection moved from sampling to every unit.
Process plant
- Challenge
- Downtime was discovered, never predicted.
- Solution
- Vibration and thermal signals modelled against the maintenance history.
- Outcome
- Work orders started following the signal instead of the calendar.
Multi-site manufacturer
- Challenge
- Each plant reported scrap and OEE its own way.
- Solution
- One analytics layer fed from the existing SCADA and MES.
- Outcome
- Plant and group argued about decisions instead of numbers.
Teams we work with
Who we build for
Different teams, different first project. The platform underneath is the same.
Quality teams
Every unit inspected, defect classes you define.
Maintenance and reliability
Vibration and thermal signals, not a calendar.
Plant management
OEE, scrap and downtime across lines and sites.
Plant IT and OT
Edge hardware, patched and monitored.
Scope
What we deliver
Four things, in this order. Each one is useful on its own.
Line-side discovery
A week on the floor. The problem worth solving first.
Vision inspection
Cameras, lighting, model and operator interface.
Edge inference stack
Local compute with safe degradation and auto-sync.
Plant analytics
Scrap, throughput and downtime both sides trust.
Engagement
How the engagement runs
No surprises about sequence, and no invoice before there is something to look at.
-
Floor walk and baseline
Measured before anything is proposed
-
Proof on one line
One defect class, real product
-
Harden for production
Failure modes and operator training
-
Replicate across lines
The tenth is a rollout, not a project
Under the hood
What it is built on.
Your existing systems stay. We add the layers that are missing and run them.
Plant systems
- PLCs over OPC UA
- MES and SCADA
- Existing line cameras
- CMMS work orders
AI layer
- Vision inspection models
- Vibration and thermal anomaly detection
- Operator labelling workflow
- Retraining pipeline
Infrastructure
- Industrial edge compute
- MQTT store-and-forward
- Offline-safe queuing
- Remote patching and monitoring
Private cloud
Compute at the line, managed from outside it.
Edge nodes on the plant floor do the inference and a dedicated cloud tenancy carries the reporting layer. Neither one depends on the other staying up.
Explore private cloud- Industrial edge nodes
- Offline-safe operation
- Single-tenant reporting layer
- Remotely patched and monitored
What changes
- 100%
- Units inspected, not sampled
- Edge
- Inference local to the line
- Offline-safe
- Survives network loss
Worked examples
The write-ups behind this work.
Full situation, architecture and results on each one.
Predictive maintenance
Predicting failures on assets with no reliable network
38 Stations, one deployment
Document automation
Forty thousand scanned documents a month, keyed in by hand
11 → 3 People on data entry
Infrastructure modernisation
Modernising an estate nobody had a complete map of
400 Hosts mapped, none by hand
Free, no call required
The Manufacturing AI readiness checklist
The questions we work through before recommending anything — data readiness, hosting constraints, review process and the running cost at year two. Use it with any vendor, including the ones that are not us.
Get the checklist- Which inspection is worth automating first
- What the line does when the network stops
- Who retrains the model in a year
- What edge hardware costs per line
Questions
Asked first, every time
Short answers. Longer ones are a conversation.
Do we need to replace our cameras?
Often not. We assess what is installed during the floor walk. Sometimes lighting is the real problem and the camera is fine. Where a sensor genuinely cannot support the inspection, we say so and specify what would.
What happens when the network goes down?
Inference runs on local hardware, so inspection continues. Results queue locally and sync when connectivity returns. Nothing on the line waits for the cloud.
How many defect samples do you need?
Fewer than most expect for obvious defects, more for subtle ones. We give a realistic sample requirement after seeing the defect classes, and can start with synthetic augmentation where real defect images are scarce.
Can our team retrain it?
Yes, and that is the point. The labelling and retraining workflow is part of handover, because a system only your vendor can update is a system that goes stale.
Will this integrate with our MES or SCADA?
Yes, usually via OPC UA, MQTT or a direct database write. Scoped during discovery.
Vocabulary
Terms worth knowing before the first call.
The words that come up most in Manufacturing conversations, in plain English.
Tell us which line costs you the most.
Scrap rate, downtime hours or inspection labour — start with whichever number hurts.
- A reply within one working day, from an engineer rather than an account manager.
- An honest read on whether this is worth doing now, or in a year.
- No marketing list. Your details reach the solutions team and stop there.
Keep looking