A personal AI assistant that runs without internet on the client's laptop. Designed and built from scratch: from architecture to interface.
A fully autonomous AI system with three-layer memory, living on the client's machine. No cloud. No subscriptions. No risk of a data leak. An assistant that knows the business by heart — and says nothing about anything it knows.
The CEO of a jewellery company works with closed commercial information — contracts, prices, correspondence with suppliers. Public AI services were out of the question: the data must not leave the device.
An offline assistant with its own memory of the business, a translator, a document analyser, an audio transcriber and voice output. All models are proprietary and local. No internet required.
The work involves supplier contracts, purchase prices, client correspondence, gemstone certificates. None of it can go out to external AI services: that breaks agreements and creates legal risk.
Clients are in seven countries. Negotiations, documents and voice messages arrive in different languages. Public translators are barred by the same security policy.
The CEO works on the move: flights, trade fairs, negotiations in places with no reliable internet. Cloud AI is useless there — the assistant has to work always.
No complicated setup, no servers, no subscriptions. The client is not a technical specialist. Open the shortcut and the system is running. Everything else happens inside his laptop.
“I need an assistant that knows everything about my business — and that nobody knows about except me”.
Client briefThe core module is a conversational AI with three-layer memory. The assistant knows the client's business the way an experienced employee would: the terminology, the partners, the history of negotiations, the decisions taken. It answers in text — and can speak aloud when you need to listen rather than read.
An assistant with three-layer memory.
It remembers the client's business, the history of decisions, writes letters, sums up negotiations. The only module that needs memory permanently — the rest reach for it when they have to.
Two-way translation across seven languages.
Preserving the business register and the client's own terminology. A separate switch turns on a literary mode for texts that need a literary form.
PDF, Word, spreadsheets.
The assistant reads the file, extracts the substance, prepares a summary or a detailed breakdown. Multi-page contracts collapse into a list of what to look at before signing.
Transcription of calls and voice messages.
Straight after the transcript — a structured summary with the key points. The module handles any common audio and video format.
Any answer can be read aloud.
Useful on the road, behind the wheel, whenever looking at a screen is not an option. Speech synthesis is local too — the sound does not leave the machine any more than the text does.
An ordinary AI forgets the conversation within a minute. VOC Offline remembers for good. But the point is not that it remembers — it is that memory switches on only where it is needed. Pick a question on the left and watch which layers come up for it.
History and the knowledge base come up. Who the supplier is and on what terms the business deals with him — from the permanent base. What exactly was discussed last Thursday — by smart search across thirty days of correspondence. The answer rests on a real precedent, not on a paraphrase.
Memory is not touched at all. This is a general question — the assistant answers from the model's general knowledge. This is exactly where naive implementations break: they drag the business context into every request and overfit on it.
The working layer does the job. Active tasks and open questions with deadlines highlighted. Completed items clear themselves after thirty days so the layer does not turn into a dump.
The permanent base answers. Working rules are one of thirteen categories, alongside the owner's biography, projects, people, decisions taken and industry terminology.
Separating “memory versus general knowledge” solves the problem of overfitting on context and gives precise answers in both modes. The prompt is trained to tell a business question from a general one: the first goes through memory, the second through the model's general knowledge. Without that separation, an assistant asked about planets starts rummaging through contracts.
Long-term memory: the owner's biography, projects, partners, working rules, decisions taken, industry terminology, the technical constants of the business.
13 categoriesActive tasks and open questions. Deadlines highlighted. Completed items cleared automatically. This is where the client's agenda for today lives.
Auto-cleanup 30 daysFor every request — a smart search across thirty days of correspondence. The three most relevant past discussions are mixed into the context.
Top-3 context pairsWorking with international clients is the daily context. The translator is trained on business vocabulary, preserves the tone of the original and handles both short replies and long commercial texts. A separate switch turns on a literary mode for texts that need a literary form.
Подтверждаем готовность принять партию по оговорённой цене. Просим выслать сертификаты и транспортные документы не позднее пятницы.
We confirm our readiness to accept the shipment at the agreed price. Kindly send the certificates and shipping documents no later than Friday.
Wir bestätigen unsere Bereitschaft, die Lieferung zum vereinbarten Preis abzunehmen. Bitte senden Sie die Zertifikate und die Versanddokumente spätestens bis Freitag.
Nous confirmons être prêts à réceptionner le lot au prix convenu. Merci de nous faire parvenir les certificats et les documents de transport au plus tard vendredi.
Wij bevestigen bereid te zijn de zending tegen de overeengekomen prijs af te nemen. Gelieve de certificaten en de vervoersdocumenten uiterlijk vrijdag toe te sturen.
我们确认愿意按约定价格接收该批货物。请在本周五之前寄送证书及运输单据。
Drag a file in, get a readable summary. The assistant reads multi-page contracts, commercial offers, certificates and reports, and pulls out only what matters.
A long call with a supplier, a voice message in Frisian, a video recording of a meeting — all of it turns into text and a structured summary with the key points and a conclusion. The module handles any common format.
— Parties and term of the agreement
— Key financial conditions
— Delivery dates and schedule
— Liability and penalties
— What to watch for before signing
Inside the system only proprietary, locally deployed neural models are running. Every computation, all of the memory and the voice output happen on the client's machine — not one request and not one file goes out to the internet.
An ordinary web application served on localhost. No installer, no server outside: the client opens a shortcut and the browser opens his own machine. Dark and light themes — because the work happens both on a plane and at a trade fair.
The only place where decisions are made. The orchestrator identifies the type of request, assembles context from the layers it needs, keeps the history, and in the background runs fact extraction from the conversation that has just happened — so that tomorrow the assistant knows more than today.
Three layers with different lifespans: permanent, working and search. The separation is not there to make the diagram pretty — it is what keeps the system from dragging everything it has ever learned into every request.
The modules know nothing of one another and duplicate no logic: each receives an already assembled context from the orchestrator and returns a result. Adding a sixth module means writing one service, not rewriting the system.
The bottom layer is the whole point of the exercise. The models run on the client's graphics card. Not one request, not one file and not one second of audio leaves the laptop.
A custom fact-extraction pipeline is the heart of the memory. Without it the system would be an ordinary chatbot. With it — an assistant that remembers everything ever discussed with the client.
Ten hours of continuous work. The checkpoints — every fifth window — are marked in gold. In a clean run they are not needed and simply pass by.
A failure at window 87. Without checkpoints all 87 are lost and the ten hours start again. With them, two windows are lost: the system returns to window 85 by the resume flag and carries on. That is the difference between “it works” and “it works in production”.
Ten hours of continuous work mean that any failure in the middle is a catastrophe. Solved with checkpoints every five windows plus a resume flag from the last saved point.
Validation runs record by record, not in whole batches. If the model produced invalid JSON in one fact out of ten, the nine valid ones are still saved.
If the model invents a field value outside the permitted list, the validator quietly maps it to the nearest correct one — instead of throwing away the whole record.
The prompt is trained to tell “a question about my business” from “a general question”. The first goes through memory, the second through the model's general knowledge. With no overfitting on context.
Not a demonstration of capabilities but a daily working instrument — on business trips, in negotiations and in the quiet of the office.
Works anywhere — on a plane, at a trade fair, in a hotel with no Wi-Fi.
No monthly payments for AI services. Everything is inside the laptop.
Not one line of correspondence or of a document leaves the device.
Contracts, prices and supplier correspondence no longer pass through external AI services.
A summary of a long contract — in minutes instead of hours. A draft letter in any of the seven languages — in one touch.
The system does not depend on the policies of third-party AI companies, on their prices, outages or regional restrictions.
The assistant knows the industry terminology, the names of the partners and the client's own style — not an abstract AI but “his own” specialist.
This project is about the fact that a serious AI instrument can belong to one person: run on his laptop, remember his business, keep his secrets. Not a service. Not a subscription. An instrument.