Automation & AI Engineer · Toronto, ON

Automate the work nobody wants to do.

I'm Yaroslav. For five years I've built workflow automation and AI agents for teams buried in manual steps — mapping how the work actually happens, shipping the system that removes it, then measuring what it gave back.

100+production workflows running today
620hmanual work removed per month
140+automations shipped to clients
5 yrsbuilding automation full-time
Selected work

From a business problem to a system that runs itself.

Client names are withheld where the work is under agreement; reported figures come from the owners and department leads who measured them.

Beauty salon · Toronto 13 workflows in production

Salon operations, running end to end

The problemBookings, client notes, reviews and follow-ups lived in different places. Nothing talked to anything, so the work fell on whoever remembered it that day.

What I builtA connected system rather than a set of scripts: customer sync from the booking platform, a daily pull of client notes, outbound message campaigns with cooldown rules so nobody gets contacted twice, review scanning with drafted replies, a bi-weekly staff report, plus a pipeline monitor and error handler that page me before the owner notices anything.

+73%review volume 26 / 170lapsed clients recovered, campaign still running

n8n · Square API · Google Business Profile · Sheets · Telegram alerting · scheduled orchestration

AI pipeline

Content pipeline from a single voice note

The problemPublishing one post meant transcribing a recording, writing copy, finding an image, posting it, then logging it. Five tools, every time.

What I builtA pipeline triggered from a chat message. It takes text or voice, transcribes the audio, drafts the post through an LLM agent with a tuned prompt, publishes it with a generated image, then extracts metadata into a spreadsheet and calendar. Access control, command routing and conversation memory included.

1 messagefrom voice note to published, logged post

n8n · Whisper · LLM agent · image generation · Sheets & Calendar API · Telegram Bot API

Content system3 workflows, always on

LinkedIn publishing with its own feedback loop

The problemPosting consistently is a scheduling problem. Knowing whether any of it worked is a different problem, and the analytics only exist as screenshots.

What I builtTwo scheduled publishing pipelines that research, draft and post on their own cadence — plus a stats workflow that reads screenshots of the platform's analytics through a vision model, writes the numbers to a sheet, and sends a weekly performance summary.

Weeklyperformance report, assembled without manual entry

n8n · Claude · vision model · Google Sheets · Telegram · cron orchestration

Voice agent

Booking agent on voice and chat

The problemA service business lost bookings during its busiest hours for the dumbest reason: nobody was free to answer the phone.

What I builtA voice agent and chat interface sharing one backend, wired to the booking system through MCP. It handles hours, availability and rescheduling on its own, and hands anything unusual to a person with the context already gathered.

Unattendedroutine bookings handled with no staff involved

n8n · voice AI agent · MCP integration · booking API · chat memory

Hostel · Poland

Taking the phone queue off the front desk

The problemGuests could not get through by phone, so the complaints landed in Google reviews instead. The rating sat at 3.7.

What I builtWe traced what was actually driving the bad reviews, then put a voice agent in front of the phone line to answer the questions that made up most of the call volume. Over half of callers got what they needed without ever reaching the desk.

3.7 → 4.1Google rating · ~24% more footfall

n8n · voice AI agent · telephony integration

43-person team

Email campaign production, rebuilt

The problemEvery campaign was assembled by hand: copy into templates, templates into the sender, lists rebuilt from scratch each time.

What I builtTemplated production that took the manual assembly out of the loop, so the team spent its time deciding what to send rather than how to send it.

~170hremoved per month · 30% faster cycle

n8n · templated HTML · marketing platform integration

RAG

A knowledge base of my own workflows

The problemEvery new automation started from a blank canvas, even when a near-identical one had been built months earlier and forgotten.

What I builtA searchable database of existing workflows stored as vector embeddings, queried in plain language by an agent that returns the closest working pattern and explains how to adapt it.

Reusepatterns retrieved instead of rebuilt

n8n · Supabase vector store · embeddings · RAG · AI agent

Side project

A YouTube channel that runs without me

The problemI wanted to find out how far an unattended content pipeline could actually go before a human has to step in.

What I builtA daily pipeline that writes the script, generates voiceover and visuals, assembles the video and publishes it on schedule — with a separate push workflow for uploads and a pull for analytics.

Dailypublishing on a schedule, hands off

n8n · Claude · ElevenLabs · image generation · YouTube Data API · VPS

How I work

The part that isn't code.

Most failed automations are not technical failures. They automate a process nobody actually follows, or land on a team that was never shown how to run them.

Sit with the team first

The documented process and the real one are rarely the same. I map what people actually do, workarounds included.

Build the smallest thing that solves it

Sometimes that is an AI agent. Often it is a scheduled script. The tool follows the problem, never the other way around.

Document it and hand it over

An automation only I can fix is a liability. Whoever uses it gets documentation and a walkthrough.

Agree the number, then check it

We decide what success means before I build — hours saved, errors avoided, time to answer — and measure it after.

Toolkit

What I reach for, and what I skip.

No tool here is the point. The point is picking the one that makes the solution boring enough to maintain after I hand it over.

Automation

  • n8n (self-hosted & cloud)
  • Microsoft Power Automate
  • REST APIs & webhooks
  • MCP integrations

AI

  • AI agents & assistants
  • Claude, OpenAI
  • Whisper, ElevenLabs
  • RAG & vector embeddings
  • Voice agents

Code & data

  • Python
  • JavaScript
  • SQL, MySQL
  • Supabase

Running it

  • Docker
  • VPS / self-hosted
  • Monitoring & alerting
  • Docs & handover

Got a process that eats the week?

Tell me what it looks like today and I'll tell you honestly whether it's worth automating — and what it would take.

Write in Gmail LinkedIn