Tomasz Kowalski
CTO at Sembot · Founder, Iseno

Tomasz Kowalski

AI-native software development·Software architecture·Go

10+ yrs in architecture, SRE and team leadership

I moved a SaaS company to AI-native software development — in production, not as a pilot. At Sembot (e-commerce SaaS) agents write most of the code, CI gates guard quality, and people decide what we build and when we ship. A year ago 12 people worked in our development and QA. Today it's 3 developers and 1 QA, and we ship more than ever.

Track record

A year of AI-native
delivery in numbers
1,600+
merged PRs in 30 days, ~95% of commits written by agents
12 → 4
people in development & QA — 3 developers and 1 QA
30 min
median from PR open to merge (18–21 h a year ago)
5 mo
to ship the new Sembot: ~20 Go services replacing the monolith
Tests grew with it: Laravel 97 → 1,092 test files, Go 27 → 1,124, E2E scenarios 64 → 409. The new Sembot went through 860+ production releases and the legacy system was switched off on 7.10.2026.

How I can help

Talks, workshops
and advisory
01 — Talks
Conference talks
What changes when agents write the code: who is still needed, what breaks, and the numbers behind it. Straight from a production setup, including what went wrong.
02 — Workshops
Workshops for engineering teams
Hands-on sessions on Claude Code and Codex in real delivery: CI quality gates for agent-written code, worktrees, review without the bottleneck.
03 — Advisory
Rebuilding the delivery process
For companies moving from “AI assists developers” to “AI writes, humans decide” — process, roles, guardrails and architecture (Go, DDD).
Talks
liderzy.it — „Budowa oprogramowania w erze AI” — 11.2026
„Kod na żądanie, development w zarządzie. Kto jeszcze jest potrzebny?”Whole conference: 4,400 registered, ~1,000 watching live at any time, 44 speakers in 3 parallel tracks„AI is not taking our keyboards away — it is giving us back our thinking.” I was too cautious.The developer no longer builds. The developer maintains.Who is still needed? The one who decides, the one who validates, the one who guards the foundations.Code is on demand. Thinking — not yet.
Ecommerce.pl Event 6.0, Warsaw — 10.2026
Workshop „Q4 bez niespodzianek. Zbuduj system kontroli konta”Whole conference, in person: over 600 attendeesYou already have alerts. Why don't you read them? Threshold without a sample, no owner, no suppression.Four steps: event, process, task, resolution. Three of them happen without you.„A 25% drop” is a bad threshold — until you say on how much data. Change thresholds and minimum samples.Q4 starter set: five processes that make sure you are selling at all.
AgenticCoding.pl — 11.2025
„Developer w epoce AgenticAI – między wizją a walidacją”Single track: ~2,000 registered, ~800 watching onlineFrom coder to designer of intent: how it used to be, vibe coding, what the few already do.If AI writes the code… who is the developer? System architect, guardian of test quality, guardian of code consistency, filler of test gaps.Feedback, not the prompt: the greatest value of AgenticAI lies in the validation loop.„AI is not taking our keyboards away — it is giving us back our thinking.” AgenticAI won't replace developers, but developers who work with it will replace those who don't.
Courses & certificates
DataWorkshop: Software 3.0 – Agentic AI Systems (levels 4 & 5)
2026
DataWorkshop: Practical LLM
2026
AI_devs 3: hands-on Generative AI course
2024
AI_devs 2: integrating GPT-4 with applications and automation
2023
DataWorkshop: NLP Challenge (editions 4 and 5)
2019
Professional Scrum · Attacking and defending web applications (Niebezpiecznik)
2016 – 2018

How we build

Agents write.
People decide.
Orchestrated agents
Hermes, an orchestrator agent I built on Codex, triages tasks every few hours, runs Claude Code sessions in isolated worktrees, spins up a 24-hour test environment and hands a draft PR with screenshots to QA.
Quality gates, not manual review
A test for every change (a change without a test fails CI), technical debt that can only go down, quarantine for flaky E2E tests, machine-by-machine rollouts that stop on the first error, monitoring with escalation.
QA and business ship
Most of the code is now shipped by QA, the C-level and Customer Success working with agents. Developers guard patterns, performance and security.
Architecture that fits AI
A Go gateway and ~20 Go services that replaced a Laravel/Angular monolith feature by feature (strangler pattern), in 5 months.
Stack & specialisms
Claude Code Codex Agentic AI Go DDD / CQRS CI/CD GCP SRE Playwright

Experience

Architecture, SRE
and leading teams
CTO · Sembot
06.2025 – present
SaaS platform for e-commerce and agencies: campaign, product feed and ad account automation (Google Ads, Merchant Center, Bing Ads). Stack: Go, Laravel, Angular.
AI-native development (2025 – present)
  • Rebuilt the delivery model: from 12 people (dev + QA) to 3 developers and 1 QA, with higher throughput. Tickets from QA and Customer Success are implemented by agents, and the C-level ships product changes with agents.
  • Orchestrated agents: Hermes, an orchestrator agent I built on Codex, triages tasks every few hours, runs Claude Code sessions in isolated worktrees, spins up a 24-hour test environment and hands a draft PR with screenshots to QA.
  • Quality gates, not manual review: A test for every change (a change without a test fails CI), technical debt that can only go down, quarantine for flaky E2E tests, machine-by-machine rollouts that stop on the first error, monitoring with escalation. Result: 1,600+ PRs in 30 days, median time to merge 30 min (previously 18–21 h).
  • Tests in one year: Laravel 97 → 1,092 files, Go 27 → 1,124, E2E scenarios 64 → 409.
  • Earlier (2025): CodeGen Manager (Go): Jira task → AI agents in Docker containers → draft PR on GitHub, with a library of 45 agents in a pipeline planning → analysis → implementation → verification → auto-fix.
New Sembot – service architecture (05.2026 – 10.2026)
  • Designed and led the build of the new Sembot: a Go gateway (templ, HTMX, Alpine.js) with a single access-control layer and ~20 Go services with their own data (CRM, notifications, telephony, invoicing, public API, OAuth connection vault, AI Marketer).
  • Gradual takeover of the Laravel/Angular monolith, switched feature by feature; legacy system switched off on 7.10.2026, about 5 months from start. 860+ production releases.
  • Q4 traffic readiness: from 22 backend calls on page load down to one, 34 h from first commit to production.
  • Built Sembot Sales Manager from scratch: a multi-tenant sales CRM (leads, sequences, mailing, Twilio telephony, reports), 430+ releases.
Architecture, DevOps and team
  • Domain architecture standard (DDD/CQRS) for the Laravel backend, with architecture tests; Google Merchant Center integrations.
  • CI/CD with container deployments (GitHub Actions, Docker, Traefik): 8 development environments, branch environments, QA, beta and production, release and rollback process.
  • Onboarded Customer Success and support to work with Claude Code (skills, documentation for agents): they ship fixes reported by customers themselves.
Stack: Go, PHP 8 / Laravel, Angular, HTMX, Claude Code, Codex, MCP, Docker, Traefik, GitHub Actions, n8n, Playwright, Cucumber, Sentry, OpenTelemetry
Senior Site Reliability Engineer · Miinto
02.2022 – 06.2025
International fashion marketplace connecting boutiques with customers across Europe. Microservices on GCP.
  • Replaced manual deployments with a fully automated process: from a GitHub Actions artifact to a multi-zone GCP rollout with database migrations, monitoring and integration tests (TeamCity, GitHub Actions, Ansible, Terraform).
  • Designed a new product catalogue system and AMQP inter-service communication libraries; fully in production after 6 months.
  • Supported the QA team technically: critical paths, pre- and post-release environment tests, protocol compliance tests; improved incident procedures.
Stack: Go, AMQP/RabbitMQ, PostgreSQL, Redis, GCP, BigQuery, Ansible, Terraform, TeamCity, GitHub Actions, Sentry, Grafana
Founder & Owner · Iseno
03.2016 – present
  • Software house: custom CRM systems and applications supporting clients' business processes. Technical sales, estimation, solution architecture, team leadership and maintenance of deployments.
Co-founder & Senior Software Architect · e-ux.pro
10.2018 – 04.2022
  • Responsible for technology: application architecture, standards, the delivery process and the development team. The agency was recognised by The Manifest (2021).
Senior Architect Developer · ENetProduction (client: Media Expert)
01.2020 – 09.2020
  • Sales system of one of the largest electronics retail chains in Poland: cart, product availability, checkout, API integrations with banks (purchase financing).
Architect Developer · IncredBots
05.2019 – 12.2019
  • Architect at the software house delivering the Evenea event platform (Symfony, React).
Architect Developer · Evenea
05.2019 – 12.2019
  • Architecture and rollout of the new version of the Evenea event management and attendee registration platform (Symfony, React).
Earlier positions (2014 – 2019)
Senior Backend Developer · IT Safety
10.2018 – 04.2019
  • Development of a web platform in PHP 7 / Symfony and MySQL. Reference from the CEO.
Senior Backend Developer · PMQ
04.2018 – 09.2018
  • Design and development of LTTS software for a private healthcare facility (2 projects) and its rollout at the client (PHP, Symfony, PostgreSQL), server administration. Reference from the board.
Senior Software Developer · GoldenSubmarine
03.2015 – 03.2018
  • Design and development of web portals for agency clients (PHP, JavaScript, MySQL), server administration. Leading complex projects from analysis to rollout. Reference from a board member.
Software Developer · Jaskółka Catering
09.2014 – 12.2015
  • Custom backend of a catering company's online shop, new features and technology advice.

Skills

What I work
with every day
AI in development
Claude Code (subagents, worktrees, skills), Codex, agent orchestration, MCP (Figma, Playwright), CI gates for AI-written code, LLM-as-judge, FinOps for LLMs
Architecture
DDD, CQRS, microservices, strangler pattern, event-driven (AMQP/RabbitMQ), API design
Backend / Frontend
Go, PHP (Laravel, Symfony), Python · Angular, React, templ, HTMX, Alpine.js
Cloud / DevOps
GCP, BigQuery, Docker, Traefik, Terraform, Ansible, GitHub Actions, GitLab CI, TeamCity, n8n
Quality
Playwright, Cucumber (BDD), PHPUnit, E2E and contract tests · Sentry, Grafana, OpenTelemetry
Management
rebuilding a team's way of working around AI, onboarding non-technical teams to work with agents
Education
Polish Naval Academy, Gdynia
2012 – 2016 · Engineer, Internet technologies
Languages
Polish
native
Spanish
fluent (C1)
English
good (B2)
Outside work: swimming coach (Total Immersion), diving (PADI OWD), IoT

Contact

Let's talk about
your delivery process
Talks, workshops and advisory on AI-native software development.
[email protected] →
liderzy.it — „Budowa oprogramowania w erze AI” — 11.2026

„Kod na żądanie, development w zarządzie. Kto jeszcze jest potrzebny?”

A year ago I told a conference audience how Sembot hands code writing over to AI agents. Today I know I was too cautious. Code has become an on-demand resource — like a subscription. Product development is now run by technical C-level executives working with agents, together with the support and maintenance team, who ship features and fixes themselves. No marketing: the process from decision to production, why we invest in QA instead of more programmers, how many agents work in parallel and what it cost us — including the dark side. And finally, the question from the title.

Whole conference: 4,400 registered, ~1,000 watching live at any time, 44 speakers in 3 parallel tracks

Slides from the talk (in Polish)
Ecommerce.pl Event 6.0, Warsaw — 10.2026

Workshop „Q4 bez niespodzianek. Zbuduj system kontroli konta”

Q4 is thirteen weeks and several peaks, not a single Black Friday. In this workshop we built an ad account control system live: instead of more alerts nobody reads — processes with a threshold, a minimum sample, an owner and suppression, which turn an event into a task on their own. Two live blocks: safety (broken landing pages, broken tracking, accidental pauses) and money (budget pacing, disapproved bestsellers in Merchant Center). Participants left with a starter set of five processes and a 48-page guide.

Whole conference, in person: over 600 attendees

Slides from the talk (in Polish)
AgenticCoding.pl — 11.2025

„Developer w epoce AgenticAI – między wizją a walidacją”

AgenticAI is not „vibe coding” but a new model of collaboration: code no longer starts at the keyboard, it starts with intent. The developer stops being the one who types and becomes the validator of autonomous agents' work — the architect and the guardian of test quality and code consistency. I walked through the full flow from a Jira task to a release with an agent in the middle, and why the real value lies in the validation loop: feedback, not the prompt. Closing demo: from intent to pull request.

Single track: ~2,000 registered, ~800 watching online

Slides from the talk (in Polish)