I'm Vitaly Kirkpatrick. I build AI agents and automations that do real business work, and everything on this site is a working demo, not a promise.
I'm Vitaly Kirkpatrick, an AI automation builder. I make AI agents and automations for business, I teach near-infrared spectroscopy, and I write memoir stories. Flip the cards below to read how six of my builds work, or press Try a demo for the short story of how I got to full AI autonomy.
Click or tap a card to flip it. Press Learn more for the longer write-up, or open the full work page.
Agent fleet
The agent fleet is a team of AI agents that do business work on their own. One agent plans a task, another writes the code or the document, another checks it, and each one leaves a record of what it did. It solves the problem of one person trying to do the work of many without losing track of it. It is built from small services that call each other over internal APIs, a shared task board, and a search index over our own code and procedures.
CRM
CloviCRM is a customer system where AI agents do more than suggest. When an agent decides a follow-up is needed, it can actually send the task, update the record or schedule the next step, and the action is logged. It solves the problem of a CRM that is only a place to type notes. The product is a web app with sign-in, built so each company keeps its own data. A seeded demo account with fake customers is needed before it can be shown.
Spectra Lab NIR explorer
Spectra Lab is a workspace for near-infrared spectroscopy, the method that reads how light passes through a sample to tell you what is inside. It offers a calibration check, guided journeys for people new to the method, and a partial least squares lab that runs in the browser. It solves the problem that NIR tools are built for experts and hard to learn. The public page is live; the full app is admin-only for now.
Animated lessons
Animated lessons turn a plain slide into a short drawn animation. Instead of bullet points, a lesson shows objects in a simple isometric style: a sample, a beam of light, a chart that grows. It solves the problem of dry technical teaching that people stop watching. The method is a pipeline that takes a lesson script, draws the scenes, animates them and renders video. A pilot of four slides exists, and no public page is available yet.
Charts engine
CloviCharts is a chart service. You send it data and a chart type, and it returns a finished chart, with more than 24 types available. It solves the problem of every team rebuilding the same bar, line and scatter charts by hand, and of reports where the charts look different on every page. It is an API, so other tools, including our own reports and dashboards, can call it. The site is public and live, so you can try it.
AI trading bots board
IntelliTrader is a board for AI-assisted crypto trading and trading bots. It shows markets, strategies and risk limits in one place, so a person can see what a bot would do before it does anything. It solves the problem of bots that are black boxes. Today it is simulated: the site says live trading and bots are planned. There are no performance claims here, and none should be read into the screens.
My videos and the VitalyBook channel: stories from my life told as drawn animation, and short lessons on near-infrared spectroscopy.
Watch the channelFour short animated explainers. They play when you scroll to them. Press Play to watch again, or tap a numbered dot to jump to a step.
Press Play, or tap a step.
Every job here follows the same road, so nothing gets lost and nothing surprises you. It starts with a request in plain words. The request becomes a ticket, a written record with an owner and a status. Next comes a plan that lists the steps, the order and the likely cost. Most plans move straight on. Two kinds never do. Anything that spends money, and anything that sends an email to a customer, stops at a human gate until a person approves it. Once the gate opens, an agent does the work. Then a separate evidence check looks for proof, such as a saved file, a screenshot or a passing test, before the ticket is allowed to close. The ticket keeps the whole trail, so you can see who asked, what was planned, who approved and what was delivered. That is how automation stays fast without becoming reckless: speed for the routine parts, and a human hand on the parts that matter.
Press Play, or tap a step.
Not every job needs the biggest brain in the building. One day the AI bill hit $12 because every task, easy or hard, went to the most expensive model. That is like hiring a crane to lift a coffee cup. The fix is to weigh first, then pour. Each job gets a quick look at how difficult it is. Routine work, such as sorting files, renaming things or writing a short summary, goes to small, cheap models that finish quickly and cost very little. Genuinely hard work, such as untangling a messy plan, goes to a large model that can think longer. The cost meter shows the result: the same amount of work, a small fraction of the spend. Penny the piggy bank, who once fainted at the bill, now sleeps through the day. Cheap where it can be, careful where it must be.
Press Play, or tap a step.
An AI agent can say a job is finished when it is not. It is not lying on purpose; it simply reports what it meant to do. That is why building and checking are two different jobs done by two different agents. The builder makes something and says it is done. The reviewer never relies on that claim. It opens the folder, looks for the file, runs the test or loads the page, and compares what it finds with what was promised. When something is missing, the work goes back with a plain note about what failed. The builder fixes it, the reviewer looks again, and only then does it pass. This costs a few extra seconds and saves a lot of embarrassment. Trust comes from evidence you can see, not from a cheerful message. In this setup, a claimed OK does not count until someone has checked it.
Press Play, or tap a step.
A website here is built in a fixed order, and each stage leaves something you can inspect. Research comes first: what people search for and what they need to understand. The plan turns that into a list of pages and the job each page must do. Then the pages are drawn as simple panels, before any code is written, because changing a drawing is cheap. Screenshots are taken at phone width and at desktop width, since most visitors arrive on a phone. Anything that looks wrong gets fixed, such as clipped text, crowded buttons or odd gaps, and then new screenshots are taken to confirm the fix. All of it is gathered on a private review page, where the owner can approve each part or ask for changes. Publishing happens last, and only after a clear yes. Until then, nothing is public.
AI automation for small business means taking the jobs that repeat every week and letting software do them. Think of answering the same customer questions, sorting incoming leads, writing first drafts of quotes, updating records, and building the weekly report. A person still decides what matters. The software does the repeated typing, copying and checking, and it never asks for a long lunch.
Start with one job that costs you an hour a day. Write down the steps. If a new employee could follow them with a checklist, an automation can too. Build that one first, watch it for two weeks, then add the next. Small steps keep the cost low and the risk small.
I build AI automations and the agents that run them: workflows that connect your email, CRM, spreadsheets and documents, with a human approval step wherever a mistake would be expensive. The proof today is my own platforms, shown in the demos above. I do not have client case studies yet, and I would rather say so than invent them.
An AI agent is software that is given a goal, picks the steps, uses tools, and checks its own result. A chatbot answers a question. An agent finishes a task: it searches, drafts, files, sends for approval, and reports back. For a business, that difference is the point. A chatbot talks about the work; an agent does it, which is more than some meetings manage.
One large agent that does everything is hard to trust. A team of small agents, each with one job and a record of what it did, is easier to test and easier to fix. My own agent fleet works this way: a planner, a builder, a checker and a reviewer, with a task board that shows who did what.
Good agents know their limits. Safe changes can be approved by a decision layer. Risky ones, such as spending money or sending a message to a customer, go to a person first. That rule is built in from the start, not added later. If you are exploring AI agents for your business, ask any builder how it works, and ask to see the log of what an agent actually did.
Have work that repeats every week, or just a question? Send me a note.