Seeing a recurring defect before your customer does.
Manufacturing does not usually lose money in one dramatic failure. It loses it in deviations recorded five different ways, corrective actions chased by email, and a pattern that only becomes visible once a customer has already complained about it.
A printed carton and label manufacturer
Around three million pounds in turnover, serving food, cosmetics and consumer goods brands.
Deviation reports, photographs, batch records and complaints sat across email, spreadsheets and production folders. The quality manager spent too long consolidating evidence and chasing corrective action updates, and recurring issues surfaced as customer concerns rather than internal findings.
A deviation, photograph or complaint enters one intake. The model classifies the type, extracts batch, product and customer, links the source evidence and flags a possible recurring pattern. The quality manager confirms classification, determines root cause and assigns the action.
A delivered Orchyn build. Clients are anonymised at their request, so the sector, scale and measured outcome are real and the name is withheld.
Where the time usually goes
Deviations described five different ways
Trend detection is impossible when the same defect is recorded as a different thing each time. Agreeing a taxonomy did more work here than any modelling did.
Evidence scattered at the moment you need it
Batch records, photographs and correspondence live in different places, so building the picture for a customer or an auditor is a search rather than a lookup.
Corrective actions chased by email
Owners, due dates and closure sit in an inbox, so status is whatever the last person remembered to reply about.
Supplier certificates expiring quietly
Nobody notices an expired certificate until a retailer asks for it. In a drinks producer with more than forty suppliers, that administration alone was twelve hours a week.
Where the system stops
It does not determine root cause, decide product release or hold, agree a customer remedy or close a corrective action. A suggested pattern is a prompt to investigate, never a conclusion. Batch, production and source evidence is preserved for human review, and access is limited by role and by customer because quality events are commercially sensitive.
This is the house position across every Orchyn build: AI assisted, never AI executed. A model is used where judgement over messy input is needed. Every figure is computed by deterministic code against a named source, and a person holds authority over anything with commercial or legal consequence.
Common questions
01Would this decide whether to release product?
No. That decision stays with your quality manager, and it is a design boundary rather than a preference. The system classifies, links evidence and flags patterns worth investigating. Determining root cause, deciding release or hold, agreeing a customer remedy and closing a corrective action are all human decisions recorded against evidence.
02Will pattern flags produce false positives?
Yes, and that is the correct trade. A flag that turns out to be coincidence costs an investigation. A missed recurrence costs a customer relationship. The system is tuned to surface rather than suppress, and it shows you why it flagged something so you can dismiss it quickly.
03Where would our quality data be processed?
On your own servers, inside your network. In a standard Orchyn build nothing is sent to us, to a model provider or to a data centre in another country. Quality events are commercially sensitive and often name a customer, which is exactly the kind of information that should not be leaving your estate.
04Do we need to change our production systems first?
Usually not. Where existing systems remain valuable we design around them and improve how information moves between them. Replacement is only recommended where the current infrastructure materially restricts the outcome. What does usually need agreeing first is a consistent way of describing a deviation, because that is the intervention that makes everything after it work.