What on-device AI means (and why Velia does not store video)

When someone considers putting a camera in their mother's home, the real hesitation is rarely whether it works. It's this: who else will be watching.

The answer comes down to one technical decision that almost no brochure explains, and that determines everything else: where the image is analysed.

Two ways of doing the exact same thing

The camera sees. Something has to look at that image and decide whether there's a person on the floor. That something is an artificial intelligence program, and it can live in one of two places.

On a server. The device compresses what it sees and sends it over the internet to the company's data centre. It's analysed there, and the alert comes from there. That's how most cameras sold in a box work.

On the device itself. The phone or camera does automatic analysis with its own processor. Instead of uploading a continuous feed for analysis, it communicates a signal. This is called *on-device* or edge processing.

One way to picture it: in the first case, every image is sent to an office that reviews it for you. In the second, automatic analysis happens inside the home and communicates a signal rather than a continuous feed. With explicit consent, an ephemeral still may be used during setup or as alert context; live video travels encrypted only when an authorised person opens it.

Both can work. But what each one can promise about privacy is different — not out of goodwill, but because of where the file ends up.

What leaves and what doesn't

"It runs on the phone" does not mean no data ever leaves. It helps to separate continuous analysis, authorised context and live video:

• Continuous-analysis frames. Captured, analysed and discarded within milliseconds. They are not stored as video or uploaded for automatic analysis.
• The result of the analysis — "a person has been horizontal on the floor for forty seconds" — is text. A few bytes. That's what travels to the family's phone.
• Live video travels encrypted only when an authorised person opens that screen and is not stored by Velia.

In the server model, the image is the first thing to leave, all the time, whether anything is happening or not. That file exists somewhere even if nobody ever opens it. And anything that exists can leak, can be requested, and can be watched out of curiosity.

That's the practical difference. It isn't that one company is more trustworthy than the other: it's that one has the video and the other doesn't.

What you gain and what you pay

You gain something concrete:

• A breach cannot expose a continuous video archive that Velia never stored.
• No employee can open a dashboard and look at the living room. There is no dashboard.
• The analysis keeps working when the internet drops, because the brain is inside the house.
• And there is a precise sentence you can use at the kitchen table: "analysis happens here, and Velia does not keep a video archive of your home." Precision makes the consent conversation easier without relying on absolutes.

And you pay:

• The model has to fit and run on a small device. These are far lighter models than the ones in a data centre. You make up for it by narrowing the problem: it isn't asked to understand a scene, it's asked to locate a body and say whether it's horizontal.
• It uses battery and generates heat. That's why these systems analyse one or two frames per second at low resolution instead of watching continuous video.
• And it doesn't fix physics. Without light the camera sees nothing, wherever the model happens to run.

Six questions to ask any product

You don't need a technical background to ask them. They work on whatever you're evaluating, and on ours too.

• Where is the image analysed — on the device, or on your servers? If the answer dodges the words "on the device", it's on servers.
• Is video stored? Where, for how long, and who can open it?
• When an alert fires, what exactly gets sent? Text, a still image, a video clip?
• What keeps working if the internet goes down? Detection should; notifying a phone far away shouldn't be expected to. They're two different things and a good answer separates them.
• Can an employee view my home? Under what conditions, and is it logged?
• Is any of this in the privacy policy, or only on the sales page? What isn't in the policy isn't a commitment.

Answers that don't answer

• *"Military-grade encryption."* Encryption in transit means nobody reads the data on the way. It says nothing about what happens on arrival. If it's decrypted at the other end to be analysed, the image left the house all the same.
• *"We only access it with your permission."* That describes a policy, not an impossibility. Policies change and companies get sold.
• *"We don't share with third parties."* Says nothing about the company itself, which is the one holding the file.

Two signals that rarely lie

• If the product sells cloud storage plans, or promises "review the last thirty days", the video is on their servers. Not necessarily bad — but the first question is already answered.
• If the device stops being useful when the internet drops, the brain was never inside.

Why this is a family decision, not a technical one

It looks like an engineering question and ends up being a conversation in the kitchen. The person being cared for has every right to ask what leaves her house, and deserves an answer that fits in one sentence without a brochure.

When that answer is "everything it sees goes to a company", the home becomes a monitored place, even if nobody ever looks. When it's "the picture stays here, and only a message goes out if something happens", it's still her home.

How we do it

Velia turns an Android phone you already own into a care aid that provides best-effort signals of possible falls.

The analysis runs inside that phone. The model lives in the app; one or two low-resolution frames per second are analysed and each continuous frame is discarded after analysis. What automatically leaves is a signal: what may have happened and when. With explicit consent, an ephemeral still may be used during setup or as alert context. Live video travels encrypted only when an authorised person opens that screen and is not stored.

And the same thing we always say, because running on the phone doesn't change the physics of the sensor: at night, without light, it doesn't see. When light drops below the threshold, the app says so, instead of showing a screen claiming everything is fine.