Trust
Responsible AI
The responsible position on AI, for a product that helps decide whether medical equipment is safe to use, turned out to be not having any. There is no model in Rydya, no inference, and no AI vendor in the path of any decision. This page explains that choice, and what would have to be true before it changed.
- Quality and compliance
- Clinical engineering
- Hospital operations
- Executive and finance
What is our position on AI?
Definition
AI in Rydya
Rydya contains no artificial intelligence, machine learning or generative model of any kind. No model file, no inference, no embeddings, no vector store, and no AI vendor API anywhere in the product, the worker or the platform console.
That is a statement about the code rather than a philosophy. There is no AI dependency in any package, and the product decisions that look like they might be learned are arithmetic you could check by hand.
It is also a deliberate constraint rather than an accident of timing. Dependence on generative AI for essential workflows is a stated non-goal in our product requirements, which means shipping it into a core path would require overturning a decision rather than merely writing some code.
Why a safety product should not have a probabilistic component in the decision path
Because the failure mode of a model is a confident wrong answer, and the whole point of the gates in this product is that they cannot be confidently wrong.
Consider what Rydya is actually for. It decides whether a device has had its required tests, whether an authorised person cleared it, whether an interval has lapsed, and what an outage cost. Every one of those is a question with a correct answer that a person could verify from the record. A model would introduce a component whose answer is a probability presented as a fact, into exactly the path where a wrong answer means a device goes back to a ward.
The argument for AI here is usually framed as assistance rather than decision: it only suggests, a human still approves. That framing survives contact with a busy department for about a fortnight. A suggestion presented with confidence, arriving in a workflow with a deadline, is an answer. The approval becomes a click, and the click is what everybody points at afterwards when asked who decided.
There is a narrower argument we do accept: AI could genuinely help at the edges, transcribing a technician's voice note or reading a serial number off a photograph. Those are capture problems rather than decision problems, and a wrong answer is visibly wrong to the person standing there. None of it is built, and if it ever is, it will be optional and outside the path of anything the gates depend on.
What the things that look like AI actually are
Three features get mistaken for machine learning. All three are arithmetic, and you can read every line of it.
The asset health score is a weighted average
Ten inputs, each normalised and clamped, combined with weights your organisation can edit, scaled to a hundred and banded at fixed thresholds. Nothing is trained. It has no training set and it learns nothing. With fewer than three inputs available it returns insufficient data rather than guessing.
Replacement triggers are threshold comparisons
A device crosses a number you configured, so it appears on a list for a person to consider. The system recommends nothing and decides nothing. A recommendation is not a decision, and this is not even a recommendation.
Report narratives are templates
Deterministic text assembled from real figures, not generated prose. The same inputs produce the same words, which is what makes a report reproducible and what a generated narrative could never be.
Search is SQL
No semantic search, no embeddings, no vector database. It finds what you typed rather than what a model thought you meant, which is occasionally less magical and always explicable.
What we will and will not do with AI
The boundary is not "AI good" or "AI bad". It is decision paths versus capture, and it is drawn in the product requirements rather than in a marketing position.
| Area | Position | Why |
|---|---|---|
| Return to service | Never | A gate whose answer is probabilistic is not a gate. |
| Quarantine decisions | Never | A clinical safety decision needs a person and a name. |
| Whether a test passed | Never | Your configured limit against a recorded reading. Arithmetic. |
| Compliance figures | Never | Missing data must read as missing, not as an inference. |
| Financial impact | Never | Money is measured or unpriced. Never estimated by a model. |
| The audit trail | Never | A generated record of what happened is not a record. |
| Voice notes to text | Possible, unbuilt | A capture aid. A wrong answer is visible to the person there. |
| Reading a serial from a photo | Possible, unbuilt | Same. Optional, and never required for the workflow. |
| Drafting a narrative for a human to edit | Possible, unbuilt | Only where the human is genuinely the author. |
Where AI is excluded, and where it might one day be optional
How to check this claim rather than believe it
Ask us to name the model. It is a question with no comfortable evasion, and it works on every vendor in this category.
Ask any vendor claiming AI which model, running where, trained on what, and what happens to your data when it is called. The interesting part is not the answer, it is how quickly it arrives. A product with a real model has a real answer and usually enjoys giving it. A product with a weighted average and a marketing department produces a paragraph about proprietary algorithms.
Our answer is that there is no model, so there is nothing sent anywhere, no third party in the path, and no training on your data because there is no training. If that changes, the honest thing is to change this page first, and you should hold us to that.
The specific question worth asking us, or anyone: what does your system do when it does not know? Ours returns insufficient data and shows nothing. That is the tell. A system that always has an answer is a system that is sometimes making one up.
Why this page exists at all
Because in this market the absence of AI now needs stating. Every competitor claims it, buyers have learned to expect it, and a product that does not mention it reads as behind rather than as deliberate. We would rather be thought behind than have somebody discover, during an incident, that the reason a device went back to a ward was that a model was confident. If AI is a requirement of your procurement, Rydya does not meet it, and that is not an oversight we intend to correct.
Questions
Does Rydya use AI or machine learning?
No. There is no model, no inference, no embeddings, no vector store and no AI vendor API anywhere in the product. That is a statement about the code rather than a philosophy: no AI dependency exists in any package. Dependence on generative AI for essential workflows is also a stated non-goal in our product requirements, so shipping it into a core path would mean overturning a decision.
Is the asset health score a machine learning model?
No, despite the word model in its name. It is a weighted average of ten inputs, each normalised and clamped, with weights your organisation can edit, scaled to a hundred and banded at fixed thresholds. It has no training set and learns nothing. With fewer than three inputs available it returns insufficient data rather than guessing, which is the tell: a system that always has an answer is sometimes making one up.
Why not use AI to predict equipment failures?
Because the failure mode of a model is a confident wrong answer, and this product exists to decide whether devices are safe. The usual defence is that AI only suggests and a human approves; that survives contact with a busy department for about a fortnight, after which the approval is a click and the click is what everyone points at when asked who decided.
Would you ever add AI?
Possibly at the edges, never in a decision path. Transcribing a voice note or reading a serial number from a photograph are capture problems where a wrong answer is visibly wrong to the person standing there. None of it is built. Return to service, quarantine, whether a test passed, compliance figures, financial impact and the audit trail are permanently excluded, because a gate whose answer is probabilistic is not a gate.
See it on your equipment
Live in an afternoon, useful the same week. A person replies, usually within one working day.
Contact us