My work

Classic attendance control rests on passes and good faith: a card can be handed to a colleague, a timesheet entry can be nudged. The task was to close that gap without making people queue at a turnstile.

I built a computer vision system: an employee opens the camera at their workstation, the model finds the face and matches it against the database. So that nobody can hold a phone photo up to the lens, there is a liveness check — the system asks the person to turn their head and verifies the movement is real, not an image.

Once the check passes, a start-of-day confirmation appears on screen and the event goes into the attendance system. Recognition runs on an ordinary work laptop — no expensive hardware, and no video leaving the building. 230 employees now go through the system.

The client had a lively audience with nothing to do between posts. The usual route is to ship an app to the store, but that means making people search for it, download it and sign up. Half of them drop off at each of those steps.

I built the game right inside Telegram. You tap a button in the chat and you are already playing: no install, no sign-up — the profile and name come from the messenger. One-touch controls, a round under a minute, so people play it in a queue, on the metro, between things.

Then the part it was all built for kicks in: a leaderboard among friends, a reward for coming back each day, and invites that give a bonus to both sides. A high score goes into a chat with one button — so the game spreads on its own, with no ad spend. More than 5,000 players have gone through the app so far.

Industrial equipment is hard to sell with catalogue pictures: the client cannot see what they are paying for, and it all collapses into comparing prices in a spreadsheet. We needed to show the engineering, not a photo of a grey box.

I built a site where the transformer lives in the browser in three dimensions. As the page scrolls, the model rotates and separates into assemblies — the tank, the windings, the core, the cooling system — with a description of each part appearing alongside.

The result is less a page than a guided tour inside the product. The visitor sets the pace: scroll further and the model opens up deeper. It is optimised to run on modest laptops, not only on gaming machines.

A stream of paperwork was piling up at the company: contracts, invoices, acts, delivery notes. A secretary spent the day reading them, copying details and amounts into the system and filing them away. Mechanical work — but work where mistakes are not allowed.

I built an agent that takes a document from a scan or an inbox, recognises the text, works out the document type and extracts the fields that matter: counterparty, amounts, dates, numbers. Then it finds the right record in the database itself and writes the data where it belongs.

Uncertain cases — a poor scan, an unfamiliar counterparty, mismatched totals — are not guessed at: the agent sets them aside and shows them to a human. Every action is logged, so it is always clear where any figure in the database came from. In money that is roughly 80,000 ₽ saved every month — exactly what an employee doing the same work would cost.

The main complaint about free solutions is not price but instability: the connection drops at the worst possible moment, speed sags, and moving from Wi-Fi to mobile data means reconnecting from scratch.

I built a commercial service focused precisely on stability: several servers across locations, automatic selection of the best node, instant reconnection when the network changes, and protection against traffic leaking around the tunnel.

From the user’s side it comes down to a single button: press it and it works. All the complexity of keys, protocols and server choice is hidden inside. Plus billing, access provisioning and node monitoring, so a problem shows up before a customer writes in about it. In the first month alone the client reached 250 paying customers.

When the client base is measured in hundreds of chats, sending by hand becomes a job of its own: copy, paste, do not mix people up, do not send twice, do not run into the platform’s limits.

I built a tool that takes it over: you upload the list, pick a template with name and parameter placeholders, and hit start. Messages go out with natural pauses and within the limits, so the account does not get banned.

At the end there is a report: who received it, who did not and why. Lists can be segmented, so different client groups get different messages rather than one text for everyone.