AI that does real work inside your business.
A language model that lives on your own server. An agent that answers and books for a local business. A studio that keeps one AI character the same in every picture. Apps where people just say what they want. Four of the things I have built, picked to show the range, each with the technology named and explained in plain words.
street · 4:5
beach · 9:16
gym · 1:1Four of the things I have built, picked to show the range.
Each one a real problem, solved end to end, told the same way: the problem, what I built, how it works, with every technology named. Each case has its own page with a working demo on recorded data.
In plain terms, this is custom AI development for companies: private LLM deployment on your own hardware, AI agents for small businesses that take bookings and orders, AI-generated photos and video of one consistent character, and natural-language search and forms inside your app.
A model that belongs to the company.
A software product with tens of thousands of daily users, and every one of them wants to ask the system questions in plain words. On a rented cloud model that is a bill that grows with every user, and company data in someone else’s hands.
The company’s own language model: an open model, taught the company’s vocabulary, running on the company’s own server, wired to its documents and internal systems. People ask inside the tools they already use; the model looks things up, creates records, books slots. Nothing leaves the building.
If your business has no agent, it is not in the agent internet.
Assistants are starting to book, order and ask on people’s behalf. A café with two tables and a phone has no way to be on the other end of that call, and hiring a developer for it makes no sense.
A platform where any business builds its own agent in four steps, with no developer: what you do, your resources, your schedule, your rules. Publish, and other agents and assistants can talk to it: book a table, ask a question, place an order. The owner sees every request and keeps the last word.
A customer’s assistant asks; the business’s agent checks the schedule and answers. When a slot is taken it offers the nearest free ones instead of guessing. Two bookings play in turn: a café table and a padel court.
One face. Any scene. As many as you need.
Shops, brands and creators need pictures of one person every week, in new places and new outfits. AI image services draw a pretty face once and lose it the next time. And nothing made there is yours.
My own studio, end to end: describe a character, cast a face from fifty candidates, teach a small model to keep it, then shoot any scene from one sentence: three photo formats and five-second clips. A person approves every frame; a caption and a slot in the calendar follow.
street · 4:5beach · 9:16gym · 1:1Vera, the studio’s drawn character, stands in for the photographs. The real output is photographic; the point is the same person in every frame.
Say what you feel like. The city answers.
New in a city, free this evening, and nobody to go with. The apps that exist want you to join a group, plan a week ahead, fill a form, or already have friends.
A map of small live events with an assistant inside. Type “padel today 6pm, need 3 people” and the event card writes itself. Ask for “something quiet in the evening nearby” and it finds matches and says why each one fits. Designed to the last screen; the backend foundation is built.
Whatever the assistant guessed is marked amber, so the person always knows what was theirs and what was suggested.
Start from your situation, not from the technology.
A cloud model under a business agreement, wired into your systems, with guard rails and a budget. You pay per use and start in weeks. Right for a local business, a small team, a first process.
An open model, fine-tuned, on a server you own. A one-time cost plus electricity, and the data never leaves. Right for a product with many users, or anywhere the questions never stop.
Small steps on real tasks. One process at a time.
The same four weeks for every case above. No slide decks; by week four something of yours is running.
What it does, what it must never do, what counts as done. You approve the wording before anything is built.
Wired to your systems, tested on a set of real cases, with budgets, limits and a fallback before anyone sees it.
Behind a switch, traced end to end. We watch what people change and fix what they stumble on.
The second process reuses the first one’s pieces. Each is weeks, not a rewrite.
- The code and the models, in your repository, on your servers
- The rules and prompts: readable, editable, yours
- The test set of real cases and the expected answers
- The budgets and limits, as settings you can change
- The traces: what was asked, answered, and what it cost
- A written explanation of every piece, like these pages
Five things that are true on every page above.
The technology behind the four cases. Pick a term.
The pieces that make these cases work, each explained in two sentences. The card says which case uses it.
Every other name on the case pages is explained there, next to where it is used: case 01 · case 02 · case 03 · case 04.
A language model that runs on a server the company owns, instead of in a cloud service. Questions and documents never leave the building, and the bill does not grow with every user.
Describe a process you’d like to hand over.
An inbox, a booking flow, weekly reports, a form people hate. Two sentences are enough. I reply by email with how I would approach it and roughly how long it takes.
hello@denyskozedub.comWhat should the AI take off your plate?
Two sentences about the process are enough. I reply by email with how I would approach it. Or just write to hello@denyskozedub.com.