Most AI projects fail before the model does.
I'm a product lead in Wrocław and Warsaw. I help companies choose the AI worth building, prove that it works, and put it to work safely, mostly where money moves and the rules are strict.
On this map, height is value and distance is readiness.The land is real: the Karkonosze from the Jelenia Góra basin, an hour before dawn. Elevation from Copernicus, map data © OpenStreetMap contributors.
1 of 4
Start where the work hurts, then ask two questions: how much is it worth, and how ready are we to build it? The near peak, ready now, is where to begin. The far one is the big bet: worth more, and a longer walk.
From my note: Not every product needs AI“We can't get a brand-new Mercedes in here when we have bicycles over there.”
2 of 4
When people have to open a new app and type a prompt, adoption stalls. The agents that change how a company works run in the background: the reply is drafted before you open the email.
How much should they do on their own? Try the ladder3 of 4
Discovery, a rough prototype, a measurement, then live. A prototype that works tells you more than a polished deck. My document-extraction prototype at MHR reached 92% accuracy and cut processing from minutes to about thirty seconds, and the number, not the demo, won the budget to build it properly.
More things I've built4 of 4
Written in October 2026. I'll check them in a year.
Now
An experiment
Four rungs. Each one has to be earned.
Month end. Forty bank lines have to be matched against the ledger. An agent proposes a match for each one and says how sure it is. Choose how far up the ladder it may go.
The ladder extends the human-in-the-loop spectrum from my note Practical AI: advisory only, AI-assisted draft, fully automated.
Let the agent post more on its own and fewer lines reach a person, but a wrong match can go through unnoticed. Hold it back and nothing is posted without someone looking, at the cost of an afternoon.
The figure that lets you move between the two needs JavaScript.
Time spent checking: about 26 min
Invented data. The trade-off is not.
If you can't measure it, it's a demo.
Notes
Longer thoughts on choosing, building and adopting AI, first published on the Spyrosoft blog. Newer notes will land here first.
A value-versus-readiness way to choose AI projects, and four common ways AI programmes go wrong.
Where to look for friction, six building blocks to start from, and why the best agents work in the background.
Discovery as risk management: what it involves, what you get out of it, and why it keeps going after development starts.
Lab
Small products and prototypes, made in the evenings to find out what AI can really do. The code is on GitHub.
Working prototype
A book recommender that can't invent a book.
It learns your taste from a conversation, and sharpens it if you hand over your library mid-chat: a Goodreads export, or just screenshots of your Kindle shelf. It only recommends titles it has found in Open Library or Google Books, and every shortlist ends with an honest anti-recommendation.
TypeScript · Next.js · Claude
CodeFoundations laid
Tells you when a deal is genuinely good, not just cheaper.
Built contracts first: a frozen data model that every part is written against, an evaluation slot waiting for each LLM component, and a step-by-step plan a coding agent can follow. The contracts and the plan are done; the model comes next.
Python · Pydantic · contracts first
CodePlayable
A virtual pet that teaches children Polish–English vocabulary.
Twelve-minute sessions, four learning games, spaced repetition with Leitner boxes, ninety-odd words across three levels, and every graphic drawn in code.
Python · pygame
CodeExperiments
Structured fields out of messy paperwork.
An experiment in reading receipts, invoices and ID cards: recognise the text, work out what kind of document it is, then pull out the fields with a language model.
Python · Streamlit · Gemini · Cloud Vision
CodeStudio
Sound first, then short films. Nothing is ready to show yet; I'll link pieces here as they're finished. Until then, the map at the top of the page has a sound of its own, built by hand in Web Audio rather than with AI. Switch it on and scroll back up: the contour lines breathe with it.
About
I started as an engineer, with a degree in electronics and telecommunications, and ran a fourteen-person IT and GIS team at Wrocław's water utility. Since then I've worked between business and engineering, as an analyst, product owner and product manager. Since 2021 I've been at Spyrosoft, leading products in payments, broadcast media and accounting.
The full history is on LinkedIn
Whether it's a first use case, a pilot that won't leave the lab, or a product you want to get right, the quickest way to reach me is a message on LinkedIn.