Yolky Sense is a local AI filter for home cameras. It recognizes your people, remembers false alarms and tells you only what really matters. Over two months at a real home evening false alarms fell 10×, and when the system names someone a human corrects it almost 3× less often.
Ordinary cameras detect motion the same way for everything. Working out what matters is a job that quietly gets pushed onto you.
A cat, a branch in the wind, a courier, a stranger — an ordinary camera reacts to motion the same way. Telling important from unimportant is left to you, every single time.
The first false alarms are annoying, the next ones get ignored. Within a week people simply turn notifications off — and the camera stays silent exactly when it should have spoken.
The feed shows everything, but it doesn’t answer “what happened while I was away”. You have to dig through the archive to find out.
Each dot is one event on the cameras on an average day of a real month. Drag the divider.
Average day over 29 days, Sep 4 – Oct 2, 2026, from the live database on the Mac Mini. “Regular camera” is every recorded event; “Yolky Sense” is the notifications that were actually sent. Night is the quietest part of the day.
Every part of the system exists to subtract noise, not to add it.
Remembers false alarms in specific places — a swing, tree shadows — and gradually stops reacting to them. The longer the system runs, the quieter it gets.
Remembers family members and pets — and bothers you only when someone unfamiliar is in the frame.
Video and recognition are processed on your own hardware. By default nothing leaves the home.
Notifications arrive in Telegram, which you already use.
Neighbors, couriers, passers-by — an ordinary camera reacts to all of them the same way. Yolky Sense remembers who is who and speaks up only when the person is unfamiliar.
From Sep 3 to Nov 3, 2026 the system watched the yard around the clock on one Mac Mini. Below — the combined totals. The deep dive into month two is in the full report.
camera events. Not a demo or a mock — the live home database.
Month one’s main junk source is mostly gone. Night still holds about 2% of findings — the yard is nearly quiet then.
By silhouette already 91%. Guessing “this car usually stands here” is still below the bar — fix it or turn it off.
After the unload the Mini runs more cameras at lower load. Free memory is still thin — the next lever.
Numbers come from the live Mac Mini database, snapshot Nov 3, 2026. Periods: Sep 3–Oct 3 and Oct 3–Nov 3. Accuracy is measured on findings a human reviewed. Public reports omit people’s names and pet nicknames.
Staying quiet when all is well doesn’t mean we stop being honest with you about where your data is kept.
Video and recognition are computed on your own hardware. By default nothing leaves the home. In our home, 10 cameras work this way on a single Mac Mini.
The phone connects to your box over an encrypted HTTPS connection. There is no vendor server in between, and no video is stored along the way.
You decide what to keep and what to delete. There is no middleman who sees the archive by default.
The live interface on our own home: events, entities, zones and rules, analytics and the 3D property map. If you want to try it yourself, request a temporary demo link — names are replaced, frames are stylized, and access can be revoked at any time.
We’ll review each request individually. Access is temporary and can be revoked at any time.
Not every notification has to be a buzz in your pocket. This is where we try calmer, more playful ways to see what’s happening at home — in the corner of a Mac screen and on a tiny pixel display.
A small menu-bar app turns the MacBook notch into a live ticker for your cameras. When someone arrives, the notch comes alive and shows who it is. One click opens the event card — the frame, the camera, who’s in it — and a tap on the frame switches to the live view.
Recognized someone you know? Confirm or correct it in one tap, and the system gets smarter. Recent events are a swipe away; a do-not-disturb mode keeps it quiet when you need focus.
A notification doesn’t have to live in your phone. Yolka is a pet on a Pixoo 64×64 screen in our home: it walks, eats and sleeps, and when the gate or door camera sees a person, an animal or a car, the screen slides away and shows the visitor — calmly, silently, and only when it matters.
The scenes here loop on repeat: the events are simulated.
Common technical questions live here. If yours isn’t covered — write to us and we’ll look into it.
Yolky Sense is an intelligent video-surveillance system that doesn’t just detect motion — it tries to understand what’s happening around the home. It recognizes familiar people, analyzes events, matches observations and helps tell what really matters from everyday noise.
Most smart cameras react to isolated events: a person appeared, a car drove by, motion was detected.
Yolky Sense looks at things in context: who appeared, where they were already seen, how confident the recognition is, and whether the owner should be bothered at all.
Yolky Sense is a working independent project. The system runs with real cameras around the clock, regularly gets new features and is checked in everyday conditions.
Right now the focus is on technology and recognition quality, not on mass distribution.
Not yet. The system is developed and tested on its own infrastructure. A public version has not been released.
Primarily for owners of private homes and properties who want to understand what’s happening around them — without constantly watching cameras or drowning in notifications.
The current system runs on a Mac mini M1 with 8 GB of RAM. That hardware is enough for the project’s live configuration.
Minimum requirements for other scenarios are not defined yet.
The main working environment is macOS. Windows and Linux will be considered separately.
Yolky Sense works with IP cameras that provide an RTSP video stream. The project uses and tests TP-Link Tapo cameras.
RTSP support is an important requirement, but by itself it does not guarantee compatibility of a specific model.
Yes. The system processes multiple video streams and can match observations across cameras.
How many cameras can run at once depends on the hardware and processing settings.
Not necessarily. The current setup uses a compact Mac mini on Apple Silicon.
The performance you need depends on the number of cameras and how complex the analysis is.
The core setup is built around a local server and cameras on the home network. Some features, remote access and notifications may still need internet.
In the current setup the system uses its own infrastructure and local storage. That keeps control over the server and access to the information.
Consumption depends on the hardware and load. The current setup uses an energy-efficient Mac mini designed for continuous operation.
Yes. The system recognizes familiar people, uses previous observations and can suggest the most likely matches.
If confidence is too low, the result can be reviewed and confirmed manually.
No. Confirming new observations is part of everyday use — a few actions in the interface are enough.
As confirmed data accumulates, the system gets more information for later comparisons.
Yes. Lighting, distance, angle and image quality all affect recognition.
Instead of hiding uncertainty, Yolky Sense helps you see why a result looks doubtful — and lets you correct it.
Yes. Night observations are part of the core scenarios. Extra checks and matching across cameras help improve reliability.
The result still depends on shooting conditions.
Yes. Yolky Sense analyzes vehicles and uses extra cues, including color, to match observations more accurately.
Yes. Detection and exclusion zones are supported. They help focus analysis on the right areas and cut down on unnecessary events.
The system uses Telegram for alerts. The goal is to report meaningful events — not every bit of detected motion.
Remote access is part of the current setup. What you can do depends on the server and network configuration.
The system can run a large part of its work on its own infrastructure, but some features may depend on external services and an internet connection.
It is an independent one-developer project that combines computer vision, artificial intelligence, product design and automation.
Yes. AI is used for image analysis, object recognition, result checking and working with accumulated observations.
The source code is not published yet. Whether to open selected components will be decided as the project develops.
Yes. Useful ideas and usage scenarios are welcome. The main development priority, though, is reliability of what already works and fewer false notifications.
Right now the system is developed for its own use cases. Commercial applications and partnerships can be discussed individually.
The project is open to discussing strategic partnership and investment. The form of collaboration depends on goals, resources and the intended direction of development.
News, experiment results and system reports are published in Yolky Sense materials. View the project progress report.
Please check the FAQ first for technical questions. Feature ideas, partnership and special cases — through the form.