Predictive maintenance fails on the floor far more often than in the model. Maintenance teams have been trained by a decade of false alarms to ignore automated warnings, and a new system inherits that scepticism on day one.
All Manufacturing →
Predicting failure requires examples of failure, labelled with what failed and when. Most plants have abundant sensor history and poor failure labelling: the maintenance record says “bearing replaced” with a date that is the date of the work order, not the date of the failure, and often not the date the machine actually started degrading.
Reconciling maintenance records against sensor history to produce usable labels is typically the largest single task in these projects, and it is where a language model contributes most — reading free-text work orders, technician notes and shift logs, and aligning them to equipment and time. That is unglamorous, and it is the step that determines whether anything downstream is possible.
The second constraint is that critical equipment fails rarely, by design. A machine with three failures in five years does not support a model that learns failure signatures. In those cases the honest answer is anomaly detection against normal operation plus physics-informed thresholds, not a supervised model pretending to predict from three examples.
On the plant network, in almost every case. Operational technology networks are commonly isolated from corporate IT and from the internet as a matter of engineering practice, which makes this a natural on-premise or air-gapped deployment rather than a cloud one.
That is not a limitation. Sensor inference is well within the reach of small models on modest hardware, latency to the floor is better, and the system keeps working when the site loses its connection — which for a plant is a functional requirement rather than a preference.
abundant, poorly labelled
from work orders and notes
or anomaly detection if rare
what the team can act on
in the CMMS they already use
confirmed, clear, or ignored
Enough to cover several failures of the type you want to predict, across seasonal and load variation. Where you have the history but not the labels — which is the common case — the first phase is reconstructing labels from maintenance records, and that is worth doing even if the modelling never happens.
Then supervised failure prediction is not available and claiming otherwise would be dishonest. Anomaly detection against a well-characterised normal, combined with physics-informed limits, is the right approach — and it is a different and more modest promise, which is exactly why it should be described accurately at the outset.
It has to. An alert that does not become a work order in the system technicians already use does not become work. That integration is usually where the realistic timeline is set.
Ranked to capacity, with the evidence attached. That is what gets an alert investigated.
30-minute technical call · no deck
What we built, what stage it is at, and a link to the live thing wherever there is one.
yourOS is an operating system with an AI at its heart. You talk to it like a person, out loud or typed. It makes your calls, plans your days and works on your projects with a crew of agents, on your own private server. Our own product, built from the first commit; the waitlist is open.
Website, backend and app assets for Sonora, a free AI sound-therapy app on iOS and Android published by Cognitive AI.
Brand, website, prototype, apps on five platforms, the AI backend and the algorithm-intelligence pipeline for Podswap, a social growth engine that runs without a human in the loop.
In closed beta: Roxy, a private AI assistant with permanent memory running on infrastructure the user owns. Brand, site, product, backend and conversational design.
Live at coherence.com: brand identity generated with AI agents we directed, the motion system and website for the group behind Sonora, AboutAI, WakeUp and MudShadows. Agent work remains in build.
Live at domainbooth.com for DomainBooth, a premium domain brokerage: a new animated mark, every page redesigned and rebuilt with designed mobile layouts, and a custom CMS for its sales record, featured names, press and enquiries.
Apps, platforms and campaigns for these teams, long before the AI work. Every one is written up on the work page — what we did, and where our part ended.
