KAUFMAN
LAB.

Kaufman Lab is an AI architecture lab run by Roman Kaufman. Its subject is not software but the method by which AI agents do the work of a whole role: an editor, an analyst, a sales manager. Every system shows its volume of work, its spending and the market price of the same work. The systems together do work worth $28,400 a month at Upwork freelance rates.

Roman Kaufman is a tech entrepreneur and the founder of Kaufman Lab. He has two degrees, in engineering and in law, hence his liking for math and for things that can be verified. He does not write code by hand. He runs AI agents as a product owner: he sets the task, marks the decision points and accepts the result against thresholds announced before the run. He built all seven of the lab's systems, and each one publishes its volume of work, its spending and the market price of the same work. He gives expert commentary to Kommersant, Forbes and Vedomosti. His main experiment is this site itself: how a new domain makes its way to the top of search and into the answers of AI assistants.

The architecture of an AI agent is the order in which it works, not the language it is written in. It defines the steps of the task, where the agent decides on its own and where it stops to ask a person, what data goes in, and by what test a result is accepted or rejected. Code is only a record of that order in the language of today's generation of models. When the next generation arrives, the record changes and the order stays. That is why each of the seven projects publishes the order itself: a flowchart, the decision points and the run data. In the niche search project, the part of the flowchart that estimates demand was backtested on 3,820,805 forecasts.

Kaufman Lab's systems handle seven tasks, each of which a company usually gives to a separate person. Analytics: a review of 30 competitor websites and a search for free niches across 82,560 search phrases. Content: a pipeline that takes an article from topic to a finished set; 121 topics went through it, and 82 reached the finished stage. Agents: an assistant that takes customer inquiries in a messenger, tested on 47 messages, and a personal assistant that keeps daily tasks, 8,500 characters a day. Websites: a 19-page site that is built from data and checked before every release, and a daily check of its indexing in Yandex, Google and Bing. The list is open, and more areas will be added over time.

A development studio takes an order and builds a system for it: the client buys work on a specific task. Kaufman Lab publishes the design and its price instead. Each of the seven projects in the registry carries four values: volume of work, spending, the market price of the same work and a verdict. In total that is $28,400 a month at market rates against $77 in spending, and all of that spending belongs to one project; the other six spend nothing. The development behind it is real: 87,226 lines of code and 2,123 automated tests. Each article also reverses the usual order: "Limits" comes before "Result", so you see what the design costs before you see its numbers.

Kaufman Lab's seven systems do work that seven freelancers would bill $28,400 a month for, while the systems themselves spend $77. The market price of each role is the midpoint of Upwork's median hourly range for that role × 152 hours a month, with rates checked on October 7, 2026: web developer $4,900, market research analyst $7,200, content writer $4,200, SEO expert $3,800 (two projects), virtual assistant $2,300, customer service representative $2,200. The comparison is role against role, not system against subscription. All of the spending comes from one project of the seven, Higgsfield credits in the content pipeline at about $0.67 per article, and the author's time is not part of the $77.

A project's cost is counted in two currencies that are never mixed: money and time. Money covers APIs, subscriptions and infrastructure, in the share that belongs to the project. A Claude subscription is not part of any project's cost, because one subscription serves the whole lab. The author's time is not converted into money at all: any hourly rate would be arbitrary and would make the number impossible to check. So "$0" in the registry means zero money spent, not zero effort, and six of the seven projects are like that. The seventh spends $77 a month. The full rules are on the method page.

Every number in Kaufman Lab's registry has a source and a measurement date, and a ratio like "N times cheaper" appears only where a source exists. The best test is an inconvenient number. The lab measures its own visibility in AI answers with 24 fixed questions to AI assistants. The first run, on August 26, 2026, covered three assistants: ChatGPT, Claude and Perplexity. It got 70 answers, found 0 mentions of the lab and was published without caveats. The second run, on September 9, 2026, covered the same three: 71 answers, again zero. Yandex's generative answer was not recorded in either run. New runs have been paused since September 29, 2026.

Kaufman Lab's systems are designed to be rebuilt, and the code will be your own. The article on each of the seven projects breaks down the design: the architecture as a flowchart, the decision points and the run data, such as 121 topics in the content pipeline and 82,560 phrases in the niche search. Any modern AI model can build a system from that blueprint. The site has no ready-made prompts, and someone else's code would help little: it is a snapshot of one generation of models and goes out of date with the next, while the blueprint carries over. You pay with your own resources: data, time and a model subscription. In return, nothing ties you to the lab: no license, no subscription, no contractor you cannot replace.

Kaufman Lab publishes seven projects with their volume of work, spending and the market price of the same work, $28,400 a month for all seven, along with the cost method. The sensible order is this: find the project closest to your task in the list of projects, read its "Limits" section and its numbers, and then write to the Telegram bot with a specific question. It is the lab's only contact. The lab does not sell boxed products, subscriptions or paid access: there are no forms, no user accounts and no payments here.