The AI That Teaches Cooking

An AI that learns a master's taste and teaches it on-site — where we are now, and how it works.

Hard technical terms are marked with a red *. Tap one to jump to “Tech Terms, in Plain Words” at the bottom; use Back to return to where you were.

On This Page

Current Progress
Vision & Technical Moat
Beyond the Master, in Teaching
Technology Protected by Patent
Built to Last, Not a Prototype
Tech Terms, in Plain Words

01 Current Progress

The cooking-transfer AI we are building is not a 'someday' concept — it is a system we are stacking up, stage by stage, right now. The hardest core technology is already complete, and on top of it we have stood up the system's skeleton. This is not a demo build or a prototype made to impress; it is deep, serious engineering aimed at making it truly work in the field.

Role select — master / learner — The app, actually running Tap a screen to enlarge · swipe to see more

So far, we have reached about 80% of the complete version. The remaining path continues in the stages below.

Already complete

On-device vision AI

AI that finds objects in the cooking scene in real time (detector) and follows them without losing track (tracker) — development completed all the way through quantization and dedicated NPU acceleration so it runs fast inside the phone.

System skeleton

The master and learner apps, the backend server, and the operator's admin system are built out in detail, wired to work organically along the actual AI-teaching flow.

Privacy by design

A master's original footage is held only temporarily — encrypted — in the cloud for training, then automatically deleted when training is done. The original is never permanently stored anywhere; the data flow keeps only what learning requires.

Cloud learning & teaching link

We connected the pipeline so a big-tech, high-intelligence general AI learns a master's taste; a general AI fine-tuned on a specific dish then teaches through Galaxy XR — seeing the learner's live field of view alongside them and guiding by conversation.

Real-time role assignment

For moments where response speed is decisive, a lightweight AI embedded in the device takes over — a system where the cloud's deep learning and the device's instant response split the work and run together.

In progress

On-device + cloud AI integration

Right now we are integrating it all so that, on Galaxy XR, the detector and tracker AI work as one with the cloud's high-intelligence general AI and the lightweight AI embedded in the device — through to the AR overlay that lays guidance precisely over the learner's view.

Learning from master footage

Systematizing partner masters' cooking footage and training the AI on it.

On-site device validation

Validating the full flow used by a learner (a franchisee or a franchise's head cook) in a real kitchen.

Lighter device

Galaxy XR is well-balanced enough for cooking practice and offers a good field of view, but for greater user comfort we plan a phased shift to smart glasses.

02 Vision & Technical Moat

The technologies this cooking-transfer AI requires line up with the direction the entire relevant industry is heading. That means ours is not a business rowing into a headwind, but one achievable with the wind of a vast current at its back.

The whole world is racing in the same direction

AI sharing your field of view in real time, and helping you in conversation about whatever is in front of you — this is not something only cooking-transfer needs. Healthcare, manufacturing, education and many industries want exactly the same direction.

So two things are happening at once: big tech is betting heavily on 'on-device AI' that runs instantly inside the device, and giants including Meta, Samsung and Google are racing to ship the device that shares your view — smart glasses.

An AI trained on the master coaches the franchise cook's re-learning, live, within their view (illustrative concept, in development)

What is happening in the industry right now

Shared vision — smart glasses

Meta Ray-Ban 2025: 0M+ units sold

Samsung · Google Android XR glasses, fall 2026

Apple Vision Pro shipped · glasses in prep

On-device AI — inside the device

Google Gemini Nano · real-time Live voice

Apple Apple Intelligence (on-device)

Meta·MS·Nvidia·OpenAI lightweight on-device models, one after another

Tailoring a general AI to a purpose is already industry common sense

Just as we train one master's cooking to build an 'AI dedicated to that master,' companies worldwide are refining high-intelligence general AIs like GPT, Gemini and Claude into versions specialized for their own purposes (fine-tuning).

In other words, on an already-proven method, we tune and build an AI specialized for a Korean master's taste.

AI already understands cooking deeply

Today's high-intelligence general AIs already hold vast knowledge and insight about cooking. We are not teaching cooking to an AI that knows nothing — we are adding 'this one master's recipes and taste' to an AI that already knows how to cook.

That is a fundamentally different starting point from training from scratch. Because we lay one master's grain over deep existing understanding, deep transfer is possible even with modest data.

The era of AI building software

On coding skill alone, certain AI models already outperform senior developers — and more importantly, while a human developer usually shines only within their own specialty, AI does well across every domain. Its progress, too, evolves literally 'by the day,' so fast that the industry voices concern.

Our cooking-transfer AI is likewise being built through coding-specialized AI, and in the near future, the moment we have the cooking-learning data, a finished version ready to deploy in the field will be complete.

03 Beyond the Master, in Teaching

The performance we aim for is not a level that merely 'imitates' the master. When it comes to teaching cooking, our goal is to far surpass the master.

Cooking and teaching are different skills

Let us be clear on one thing: no one can match a master's cooking itself. The only place we aim to surpass is the domain of teaching — and it is exactly that teaching AI fills, and goes beyond.

Same one-on-one — what's different

Doing well, and teaching well

Making a dish naturally with ingrained skill, and unpacking that knowledge so others can understand it, are entirely different abilities. The cooking-transfer AI senses where a learner gets confused and what they lack, and focuses on the act of teaching itself.

Tireless, for each and every learner

Even one-on-one, a person is limited by stamina and time. AI teaches each learner anytime, any number of times, always by the same standard.

Pinpointed in view, shown in numbers

A person can only say 'that pot, over there,' but AI marks that very object in the learner's field of view and shows what was left to the eye — flame intensity, simmering time, consistency — as numbers.

Down to the finest difference, kept identical in real time

Cooking is delicate enough that one or two changes shift the taste — so even learning one-on-one, it is hard to reproduce. The cooking-transfer AI remembers the master's cooking pattern and checks in real time whether the learner's actions stay identical, closing even the smallest gaps. A person catches big differences but easily misses the details; AI does not.

And the performance keeps improving

This depth of teaching grows together with the underlying high-intelligence general AI. We will adopt each more powerful fine-tunable model as it arrives to raise performance — and we expect to reach our target, teaching that far exceeds a master's, well before the so-called era of AGI, with models available long before then.

04 Technology Protected by Patent

From the angle of technical defense

The core of this cooking-transfer AI — learning a master's taste into a standard, then comparing and transferring it on-site in real time — is protected by patent.

A master's taste, as a digital standard

Learn a master's cooking from a first-person view, and a taste too subtle to put fully into words becomes a digital standard you can work with. Heat, broth richness, seasoning, fermentation, aging — Korean cooking's delicate elements turn from scattered intuition into clear references. That very conversion is protected by patent.

Compared on-site, in real time

A learner's actions are compared against the master's standard right there, with the next move pointed out within their view. Not learn-once-and-done, but the act of transfer itself made objective — and that method is what is protected.

Korean food, judged on its own terms

Elements unique to Korean cooking — ones a Western yardstick cannot measure — are quantified into one coherent framework. That very framework for judging Korean food on its own grain is also part of the patent.

05 Built to Last, Not a Prototype

The cooking-transfer AI and the system around it are not a 'for-show' prototype or a minimum viable product. So that learners can truly learn to cook, we invested heavily in the groundwork from the design stage and built it up one layer at a time.

0 principles Software principles embedded throughout the design

Unshakable — yet always changeable

Modularity

The system is split into independent parts — auth, video, security, AI. Fixing one part does not shake the others (like building a car's engine and wheels separately).

Interface abstraction

Parts connect only through agreed-upon sockets. Swap the video input from today's XR device to smart glasses later, and the rest of the app keeps working.

Extension points

Empty slots for a stronger AI or a new device are laid in at the design stage — so there's no starting from scratch when the time comes.

Parameterization

Values that may change — security strength, evaluation thresholds — are pulled out into external settings, not buried deep in code. We adjust them without touching the code.

Encapsulation

A master's core data stays inside a safe-like 'capsule'; instead of the original, only the essential feature values are handled through defined channels.

We held to these five principles for one reason: so that when a more powerful AI or a new device appears, we can swap only the necessary parts — without tearing up the core — and carry on at once. A structure ready to take in the technologies of the future.

Head-restaurant recipe security

Wrapping the recipe asset in many layers

Recipe asset

A master's recipes and taste are an asset beyond exchange. So we placed that asset at the center and wrapped it in many layers of security — the inner layers guard what's closest to the original, the outer layers stop outside threats first. As the company grows, we've left room to add top-tier security on top.

Inner → outer: five layers of defense

Held briefly, then auto-deleted
Encrypted in transit & at rest
Kept isolated within Korea
Identity, permission & access logs
Attack defense & continuous audit

Footage used for training is kept only temporarily and then deleted; the original is never permanently stored.

06 Tech Terms, in Plain Words

(AR) overlay: Placing the AI's guidance (next move, position, score) precisely over the real cooking scene.
Galaxy XR / XR: Samsung's XR (extended-reality) device — worn like glasses or goggles to overlay information onto what you see. On-site learning runs here.
On-device AI: AI that runs inside the phone or XR device itself, not on a server — so it responds instantly, even offline.
Detector: AI that finds 'what is where' on screen — e.g., the pot, the ingredients, the hands.
Tracker: AI that keeps following the same object without losing it as it moves.
Quantization: Compressing a heavy AI to run fast on small devices while preserving accuracy.
NPU acceleration: Processing on the device's dedicated AI chip (NPU), saving speed and power — practically essential for real-time on-device work.
Fine-tuning: Training a general-purpose AI further on a specific field's material to specialize it (e.g., for one master's cooking).

Why is the 'overlay' the hardest part?

While hands and ingredients move nonstop in front of you, the detector must find each object accurately → the tracker must follow it without losing it → only then can guidance be laid precisely on top. A slight delay or misalignment breaks the learning. Because all of this must run in real time inside the device — not on a server — quantization and NPU acceleration become the crux.