Portrait of Yaroslav Kostyk on a canal street at night
Yaroslav Kostyk

AI automation engineer, Czech Republic.I build agents that do real work end to end.I build agents that do real work end to end.12 of them run on my own server right now.

Hire me
Selected work

Agents that keep running without me.

Selected work Five things I built and still run. Every one of them is in daily use.

nick! · Personal AI Assistant

Multi-agent system on Telegram

Twelve agents on one VPS, each with its own scope: inbox triage, tasks, job watchers, finance summaries, media production, remote control of my Windows PC over an SSH bridge. They hand work to each other by a written protocol, and routine checks run as plain cron scripts with no model calls at all.

  • python
  • claude code
  • telethon
  • systemd
  • notion api
  • gmail api

Subtitle Translator

Learn Languages - chrome extension

Turns the subtitles of any HTML5 video player into an interactive dictionary. Hover a word and the translation appears, computed on the device by Chrome built-in AI: free, private, offline. Split subtitle lines are stitched back into real sentences first, so idioms come out right, and saved words export to CSV, Excel, Anki or PDF.

View in the Chrome Web Store
  • javascript
  • chrome extension apis
  • on-device ai
  • deepl api

Web Studio

Agent-run production line

A small studio where the production line is agents I trained and supervise. One finds the companies worth writing to, the other keeps the client table, writes the drafts and prepares the site. I set the rules, review every output and answer for the result.

  • claude code
  • html
  • css
  • javascript

IGWatch

Follower analytics in Telegram

Tracks one Instagram account: who followed, who unfollowed, who stayed mutual, with full history. A diff engine turns two snapshots into events and a push. Read only by design, it never likes, follows or marks stories as seen.

  • python 3.12
  • aiogram
  • docker
  • sqlite
  • instagrapi

Logram

Personal Telegram utility

Makes chats readable later: local speech to text for voice notes and round videos without any paid plan, fast access to earlier history, and a full chat export into a single HTML file.

  • python
  • telethon
  • whisper
  • systemd

How I work Three rules that decide almost every technical call I make.

01

Cheap by default

If a task can run as a plain script on a schedule, it runs as a script. The model is expensive, and it gets called only where understanding is actually required.

02

Checked, not claimed

Every "done" is backed by a command output, a file or a screenshot. If something could not be verified, that is said out loud instead of being rounded up.

03

Self-taught, built the hard way

HTML, CSS, JavaScript, Python and my own Linux server, all learned on my own. What I take into my hands, I get running.

Process Agents hand work to each other by a written protocol.

Close-up of the assistant chat: a numbered task list with a suggested order
Each agent has its own scope: inbox triage, tasks, job watchers, finance summaries.
Close-up of IGWatch: a list of follower events
A diff engine turns two snapshots into events and a push.
Close-up of Logram: a voice note with its transcript
Local speech to text for voice notes and round videos, without any paid plan.

About

Role
AI automation engineer
Country
Czech Republic
Stack
Python · JavaScript · Linux
Languages
English · Ukrainian · Czech · Russian
Night on Charles Bridge in Prague, a figure in a hoodie with hands at the face

Open to remote or on-site work in automation, AI agents and IT support.