I build and ship AI products, hands-on and end to end.

Principal Product Manager, AI. I'm AI-native: I choose the architecture, build the prototype on an agent harness, and write the evals that decide what ships. Then I set the team up to work the same way.

How I shipthe loop, every time
  1. Brief
    One page. The problem, who has it, and the criteria to stop.
  2. Prototype
    A working version on an agent harness, in front of customers.
  3. Eval
    A golden dataset and scorers decide what good means.
  4. Ship
    Launch, measure, then keep it or kill it.
Product leadership at
VismeLodgifyOLX GroupFlickArex Markets

AI-native

I know which tool the problem needs.

The model is rarely the hard part. These are the calls I make on every AI product, before anyone writes a prompt.

What I reach for

A single LLM call

When
The task is one step and the context fits in the prompt.
My rule
Start here. It is cheap and fast, and I only move up when it fails an eval.

A workflow with LLM steps

When
The steps are known. Code handles what must be exact, the model handles language and judgment.
My rule
Predictable quality and cost. This is where most production AI should live.

An agent on a harness

When
The path can't be known in advance and the task needs tools.
My rule
More tokens, more latency, more ways to fail. I add autonomy only where it earns its cost.

The layers I design

Product surface

What the user sees, where a human approves, and how the agent explains itself.

Harness

LLM call, workflow or agent. Claude Agent SDK, tool calling, skills.

Context

What the model knows when it answers. MCP servers, retrieval, a governed semantic layer.

Models

Which model for which step, chosen on quality, cost and latency.

Evals and observability

Golden datasets, scorers, traces and release gates. The floor everything stands on.

What I check before an agent ships

Golden datasetReal cases, refreshed from production failures.
Deterministic checksSchema, numbers, the right tool with the right arguments.
LLM-as-judgeFaithfulness, completeness and tone, against a written rubric.
Cost and latencyMeasured per successful task, so failures aren't hidden.
Release gateA regression blocks the release, and a person approves where a mistake would be expensive.

Selected work

Four things I shipped, and what happened.

Visme, 2026

A new AI product from concept to pilot in four weeks

Problem
A new AI product idea had to prove real demand before the company committed a roadmap to it.
What I did
Wrote the concept, built the working prototype myself on an agent harness, ran discovery and pilot calls, and set the kill criteria before the test started. Then ran the launch: positioning, pricing policy, landing page and Product Hunt.
Outcome
Concept to a customer pilot programme in four weeks, decided on evidence from real customers.
Claude Agent SDKMCPVercel
See Playbooks
Visme, 2026

A product team that ships AI-native

Problem
The team planned by feature list, and every design waited on a front-end handoff before it could ship.
What I did
Replaced feature lists with outcome-based planning, introduced agent-ready tickets and internal agents for routine work, and set up the designer with a coding agent on the production front-end stack.
Outcome
The designer now commits the production design system on GitHub. Planning is tied to customer outcomes, and the front-end handoff is gone.
Coding agentsGitHubInternal agents
Visme, 2026

An evals practice that gates every AI release

Problem
An AI generation pipeline worked in demos, but nothing measured quality and nothing gated a release.
What I did
Built the evaluation practice from zero: golden datasets, LLM-as-judge scorers, eval-gated releases that catch regressions, and human-in-the-loop QA. I read the traces myself to find what the dashboards hid.
Outcome
Quality is owned by the product team, and the practice is used across product, engineering and QA.
BraintrustLLM-as-judgeTraces
Lodgify, 2023 to 2026

A partner platform that hosts could rely on

Problem
Most hosts were connected to a single booking channel, and sync failures with partners caused overbookings.
What I did
Led a group of product managers across distribution and rates, and owned the partner API platform with Booking.com, Airbnb and Vrbo.
Outcome
Listings on two or more channels went from 40% to 65%, and reservation and messaging sync reached 99% uptime.
Partner APIsMarketplacesB2B SaaS

Built

Things I've built myself.

I prototype with AI coding agents. It is the fastest way I know to find out whether an idea is any good.

Playbooks by Visme

Turns a team's best deck into a reusable playbook, so anyone can generate an on-brand deck filled with their own data. I built the prototype on an agent harness and ran the launch.

Open Playbooks

Zeniks

A multi-agent audit for Airbnb listings. An orchestrator hands each part of a listing to a specialised agent, and a final agent assembles a report with clear fixes. I built it to learn orchestration and evals by doing.

Internal agents

Small agents that take routine work off the team: a ticket writer, a call transcript classifier, and a landing-page builder that marketing uses.

Eval scorers

Scorers and golden datasets that run on every release, written with the Braintrust SDK. They are how we know a change made the product better.

10+years leading product across AI, fintech, marketplaces and hospitality
4 weeksfrom concept to customer pilot for a new AI product
30%adoption of the first OpenAI feature I shipped, within six months
65%of Lodgify listings on two or more booking channels, up from 40%
Dominik Kolacz

About

A product person who builds.

I'm Dominik, a product leader based in Barcelona. Over more than ten years I moved from pricing and data into product, and now into AI. At Visme I lead the shift from a design tool to an AI content platform: I write the strategy, own the quality bar of the AI, and build the prototypes that test where we go next.

Before that I led PM groups at Lodgify, ran a vertical at OLX Group, and shipped early OpenAI-powered features at Flick. I'm technically fluent in Python, SQL, APIs and data models, which is what operating agents well takes.

Away from work I like the ocean, surf and quiet places. That is where the calm comes from.

PolishEnglishSpanishBarcelona or remote
Open to conversations about senior AI product roles, in Barcelona or remote.

Contact

Say hello.

Email and LinkedIn are the fastest ways to reach me.