The Prompt Decalogue
Universal rules for talking to LLMs that strip out ambiguity at the input stage.
I advise companies and local governments on technology, AI, marketing and communication. I take a project from the first decision, through architecture, to deployment, and I stand behind the result. The AI I build doesn't lie. I measure quality, test for safety and cut hallucinations.
Prompt engineering isn't enough. Most models work fine in a proof of concept and then fall apart in production. Deploying AI means building an architecture that stops errors before they reach the user. I run mine as the AI Quality Loop. The whole method is on display in the experiment "One Question, Split Across 91 Minds" (in Polish).
Universal rules for talking to LLMs that strip out ambiguity at the input stage.
An architecture in which independent models (Claude Opus/Fable, OpenAI GPT Sol, Kimi K) cross-check each other's output.
Version-controlled AI context that enables a precise rollback to a stable state.
Feedback loops that automatically improve answer quality based on how the answers actually perform.

My largest and most public project: my own internet-radio operating system for the M5Stack Core2 cube. Over 200,000 people have found their way to the project, and 150+ installations are already playing in the wild — for people I have never met, on three hardware revisions. The radio plays from the first minute. There are no accounts, no cloud and no telemetry. There are 218 stations, each verified by a live connection, and setup runs from your phone over a QR code. All of it free for anyone, anywhere, GPL-3.0 licensed, with full sources published next to the binary and a live repository on GitHub. The system was built in tandem with AI, on 3.45 billion tokens of work by two agents working under my direction. The result is mine to defend.

The cube has less memory than a single phone photo, yet it must juggle Wi-Fi, Bluetooth and music at once. A few battles we won together with the agents:

A 3-in-1 device designed by AI. The measurement units ("Lego-style"), the enclosure and the architecture were all generated iteratively by LLMs. Hardware cost cut from €10,000 to €20. Version 3.0 now in the works, with a touchscreen UI, OTA updates and a measurement engine frozen and locked against drift. The entire rebuild runs through multi-agent AI orchestration (dozens of agents per pass).
A brand from scratch: the name, the colors, the direction for the logo, the slogan and the "zen" philosophy everything else grew out of. Built when Polish hosting was sold on specifications — we sold calm. In 2011 I moved the company to fully remote work, years before the pandemic made that ordinary. I left in 2017; the brand stayed and still works.

A custom-built B2B service portal for electronics manufacturers. It handles RMA tickets on its own, catches duplicates (fraud prevention) and keeps SLAs on track, so no one retypes paperwork anymore.

An enterprise-grade system for Polish municipalities. 100% offline PWA, built to WCAG 2.1, 3 languages (PL/EN/DE). Deliberately framework-free, so it runs on aging municipal servers and stays easy to audit. Version 2.0 (live): an 8,000-line monolith rewritten into a modular, testable architecture, with 611 unit tests and 144 E2E "parity" tests (v1≡v2, identical output), plus white-label for more municipalities. The rewrite itself, run through multi-agent orchestration, surfaced 7 real bugs that 35 prior audits had missed.

Monitoring of Schumann resonance, solar activity, and the geomagnetic field from 6 stations worldwide — AI translates raw data into plain language. No accounts, privacy by design, and an open API made for AI agents (llms.txt, RFC 9727).
A free biorhythm app based on Dr. Sikora's method, in 8 languages, with 288 articles and no accounts at all. In an open study of the method I run a blinded N-of-1 journal (Pearson correlations, effect size) and an anonymous opt-in aggregate. I'd rather find out honestly whether the method works than simply sell it.
The biorhythm engine as a public npm package, v2.1 (MIT): 952 golden tests, validated digit by digit against the original 1983 method. Since v2.1 it ships an MCP server, so an AI agent computes the method instead of guessing it. Anyone can open the repository and recompute all 952 cases themselves.
Self-hosted, cookie-free analytics (Plausible CE) for ~20 domains. Visitor privacy plus full data ownership.
Engagement models, pricing, and retainer (strategic partner).
View →One of the first cases in Poland of AI acting as a campaign's lead strategist. Enterprise-level capability on a local budget: communication strategy, content generation, live sentiment analysis.
Designing and deploying enterprise-grade AI systems. Focus: multi-model orchestration, hallucination elimination (AI Quality Loop), token cost optimization.
I took a manufacturing company through a full digital transformation. Online reach grew by 200%, and the M2 Service Portal I rolled out cut support costs by 90%.
Co-founder and CMO of one of Poland's first cloud hosting providers. I owned the brand, communication and growth for six years, until I left in 2017. More detail on the CASE-00 card above.
"Technology without communication is a cost. Communication without technology is a promise."
2003–2008 · Thesis: "Growth Strategies for Small and Medium Enterprises"
Strategy, audit, or implementation. Reach out through the form and I'll get back to you with specifics.
Contact form