Three kinds of help, in three parts of the hospital.
It is easier to understand what this is by looking at the work it takes off people, rather than at what is under the bonnet.
The typing, the copying and the searching that eats a clinical day.
- A doctor speaks; the consultation note is typed out and put in the right boxes
- A patient brings an outside lab report; the values are read off it into the record
- A discharge summary is drafted from what is already in the file
- The nurse going off duty gets a handover note drafted from the shift's entries
- Instructions for the patient are translated into the language they speak
- "Show me last month's diabetic admissions with HbA1c above 9" asked in plain words
Seeing what is coming: beds, stock, money and staff, before it becomes a problem.
- How full the wards are likely to be over the next fortnight
- Which medicines are heading for a stock-out, and which are heading for expiry
- Work that was done but never billed
- Shifts with nobody rostered, and leave that clashes
- A plain-English summary of the day's numbers for the management meeting
Putting what is already in the record in front of the clinician, at the moment they need it.
- A short factual briefing on the patient, pulled together from their own record
- Every line of it linked back to where in the record it came from
- Safety checks your hospital has approved, run against that patient's record
- Nothing invented, if it is not in the record, it does not appear in the briefing
The AI cannot save anything into a patient's record.
Not in any part of the system, not in any circumstance, not even when it is confident. Everything it produces appears as a draft on the screen of the person who would have written it anyway. They read it, change what needs changing, and press save themselves through the same screen, with the same checks, as if they had typed it. This is not a setting that can be switched on. The software is built so that no other path exists.
The person read it and kept it as written
They edited it first and we record that they did
They rejected it and wrote their own
All three are recorded. If a department starts accepting everything without changing anything, that shows up and we treat it as a warning sign, not a success.
What happens when a doctor dictates a note.
The most-asked-for feature, start to finish, so you can see exactly where the person sits in it.
If the feature is switched off, or the computer running it is down, the doctor types the note the way they do today. Nothing in the hospital stops working because the AI is not available every one of these features has a manual path that stays exactly where it was.
Everything goes through one checkpoint
Every request the AI handles a dictated note, a stock forecast, a bedside briefing passes through the same checkpoint before it reaches a model, and again on the way back. Nothing in the system has a private route around it.
One checkpoint, for every kind of AI work. Scroll sideways for the full picture.
Switched on for this hospital, and this user permitted to use it
Each feature declares the fields it needs; nothing else leaves the record
Text in a scanned report or a supplier's invoice is treated as text, never as instructions
Anything that reads like clinical instruction from the wrong part of the system is thrown away
What was asked, what came back, who saw it and what they did with it
How we decide what the AI is allowed to do.
Before we build any AI feature at all, it goes through the same four questions in the same order. The answer decides which rules apply to it for the rest of its life. A feature cannot quietly move between groups later that takes a signed decision, and the software checks the answer again every time the feature is used.
If none of the four questions fits, the feature is not built. There is no fifth option where we work it out as we go.
The things it does not do, said plainly.
Every one of these is a clinical decision. They belong to your doctors, and the software is built so it cannot take them.
These are the things buyers most often assume are included, so we would rather say plainly that they are not. Every one of them would legally be a medical device in India, needing its own licence and its own approvals. The software is built so that these cannot appear later by accident the code for them is kept separate, switched off, and blocked from being used by anything else.
Patient information does not leave your hospital.
Not to us, not to anyone else. The computers doing the work sit on your premises.
The AI runs on computers on your own premises. Patient information is not sent to any outside service, ours included. There is an installation path for hospitals with no internet connection at all.
Every request is tied to one hospital and one facility. Data from two hospitals can never end up in the same request, the system stops rather than allows it.
The technical records the system keeps for troubleshooting hold references and identifiers, never names, notes or results.
Dictation audio and photographed documents are kept only long enough for the clinician to check the result against them, then deleted automatically.
Rosters, attendance and payroll are personal data under India's data protection law. We treat them with the same care as patient records, and no one is subject to an employment decision made by software.
Every AI interaction leaves a record of what was asked, what came back, who saw it and what they did enough to reconstruct it later for an audit or an inquiry.
It arrives switched off, and you decide what to switch on.
A new installation has every AI feature disabled. Nothing turns itself on. Each one is enabled deliberately, feature by feature, department by department, and we record who authorised it.
Before a feature is switched on for real work, your own team tries it on your accents, your vocabulary, your forms and your patient mix. If dictation does not cope with how your doctors speak, it does not go live.
Any feature can be turned off for a department or for the whole hospital at any time, and the manual way of working is still there underneath, unchanged.
Voice is the part that needs real computing power, and it is dictation not anything else that decides the size of the machine you need. We size that from how many clinics, wards and theatres will be dictating at the same time, and we will tell you before you buy anything.
For your IT and clinical governance teams
The plain-English version above is not a simplification of a vaguer reality. It is backed by a published internal architecture document that your technical and governance people are welcome to review with us line by line.
- Every capability is assigned to one of three governed tiers, each with its own guardrail profile and intended-use statement
- All model inference passes through a single gateway; no module holds a direct client to an inference runtime
- Models are admitted to a licence-controlled registry with two-person approval, and verified by signature and checksum before load
- Assistive and operational request paths are isolated from the clinical path — no shared session, context or conversation state
- The rules above are enforced by static-analysis gates that fail the build, and runtime gates that fail the request
- Every interaction produces a hash-chained evidence record sufficient to reproduce it
Zypocare Clinical AI assists with documentation, information retrieval and operational planning. It does not diagnose, prescribe or determine treatment, and clinical decision-making remains with the treating clinician at all times. AI features are delivered in phases and are switched off by default in every new installation; what is available depends on the modules licensed, the configuration and your own acceptance testing.