Episode Materials: Leading a Team with AI with Dima Tverdokhlebov (Category #Leadership)

I’ve collected the materials from my conversation with Dima Tverdokhlebov, founder of Omnius.team (@dimatverdai), formerly a product leader at VK and head of AI at MTS. The Code of Leadership livestream took place on September 24, 2026. In the announcement I promised a conversation about moving from large companies to building your own business, but most of the time went to practice: how a nine-person agency brings generative AI into large clients without turning the founder into the sole point of coordination.

Dima himself rates his organization at three or four out of ten relative to where he wants it to be. So this is an account of work in progress, not a ready-made recipe.

We discussed:

1️⃣ Pilots on external data In Dima's experience, large clients' adoption is slowed more by the wish to keep everything inside their own perimeter and reproduce cloud model capabilities there than by technical complexity. Omnius.team starts with tasks where data flows from the outside in: the result is shown to decision makers, and only then do integration, documentation, and interfaces come up. We also noted that a successful pilot does not yet mean the core business has changed. 2️⃣ Three cases Tax legislation monitoring for an oil company, an assistant for Telegram advertising, and a search for construction material substitutes, where sixteen agents examine different constraints while separate critics challenge their conclusions. The comparison of "30–40 minutes instead of a month of work by two or three analysts" is the guest's estimate; there is no independent measurement. 3️⃣ Shared team memory Nine independent specialists are connected by an internal assistant: it sees meetings, email, and calendars, reminds people of agreements, and tasks in Linear appear as a reflection of discussions. The assistant's rules are refined through real mistakes—for example, a one-on-one conversation about salary must not end up in the shared chat. 4️⃣ Knowledge outside regenerable code If an implementation can be generated again, requirements, decision rationale, and checks should be kept separately from it. And when an agent makes a mistake, fix not only the code but also the conditions that led to it. You still need to read the code: someone has to verify the result. 5️⃣ Hiring and talant.club Dima looks for people who have already tried solving their own problems with AI, can express their thoughts clearly, and want to build useful things. At talant.club, several vetted teams propose ready solutions to one specification, the client picks one, and the winner receives half the budget. At the time of the conversation the community had five people, so it is an early experiment.

Episode materials:

📌 Episode page with timestamps 🎬 YouTube, VK Video 🎧 Podster, Yandex Music, Apple Podcasts 📝 Edited conversation recap

Dima's closing advice to managers is to solve one small problem with AI yourself, so that later decisions rest on your own experience. If you have already tried, tell us: which problem did you pick first—and what did you change in how your team works afterwards?

#CodeOfLeadership #AI4SDLC #Management #Leadership #Engineering #Product