About this site
Large AI programs fail less often on model quality than on the things around the model: how work is decomposed, how agents are observed, how cost and risk are governed, and how decisions made in a steering committee survive contact with production.
The story
The site tells one expedition. Launching and scaling an AI system is a climb. The six stages are camps on the route, each with a question the team answers before it moves on. The valleys are the failure modes where climbs stall. The radar is the gear check. The field guide is the guidebook, with sources for every stretch, and the trail log is my notes from the climb. I'm the guide: I don't carry the team, I walk the route with it.
The metaphor is a frame, not a claim. Every camp, valley and piece of gear on the site names a real decision, failure mode or practice, and links to its evidence. The home page tells the climb from the summit view down to the guide; the other pages are the reference you would carry.
The route
The route divides the life of an AI system into six stages, from the first decision to accountable operation: Purpose, Design, Launch, Scale, Operate and Govern. Each stage is a camp, defined by the question it has to answer. Govern is drawn last, but it applies from the first decision.
Notes are written for the people who make those decisions. Each one is grounded in a public engineering standard or specification, explained without jargon, and closed with requirements that can be adopted as written. The aim is guidance an executive can act on and an engineer can check.
How sources are charted
Every source in the field guide is pinned to the one or two stages where it matters most, the first being where it is drawn on the map. Each carries two lines: why it matters, for the person who decides, and what to check, for the person who builds.
- Shape
- Rule standards, specifications and regulations
- Practice frameworks, guides and books
- Evidence papers and postmortems
- Fill
- Stable settled ground
- Evolving published, still changing
- Draft not yet settled
Being specific about maturity is part of the guidance. A program built on evolving ground should plan for change rather than discover it.
The radar
The radar is the gear check, a second reading of the same ground. The route says when something matters in the life of an AI system; the radar says how far to rely on it. Each practice sits in one of four rings: Adopt is what to take on the climb, Trial what to test on the approach, Assess what to scout ahead, and Hold what to leave behind. Each is backed by sources in the field guide. The placements are editorial judgement, and each edition marks what moved.
Two voices
The route, the radar and the field guide are written to be impersonal: they are a reference, and they should read the same whoever opens them. The home page and the trail log are the opposite. The home page is my own take on building agents and AI systems, told as the climb, and the trail log keeps my dated notes on what I'm testing and what I'd watch out for. Across the site, vermilion marks my own writing and ultramarine marks the reference.
Colophon
Titles are lettered in Barlow Condensed Light, a slightly rounded condensed grotesk; plate lettering is set in Jost, a geometric typeface in the Futura line. Both are self-hosted under the SIL Open Font License. Reading text uses your system's sans serif. The pages are static HTML generated by a zero-dependency Node script from JSON files, and they read in full without JavaScript. There is no tracking and no request to any third party.
The mountains are generative. My surname, Górski, means "of the mountains" in Polish, so a small script draws each part of the site its own summit from a seed: contour lines seen from above behind the page titles and around the radar, and one ridge seen from the side that the route's six stages climb in order. They are drawn once, when the site is built, as plain images.
On the home page the same ridge becomes a panorama behind the title, and the title names its summit. As you scroll down from the title, the route draws itself up the ridge, each camp appears as the route reaches it, the trail blaze climbs to the summit and the nearer slopes settle slightly. It moves only when you scroll. On phones, with reduced motion turned on, or without JavaScript and scroll-driven animation, the route is shown fully drawn.
The site was built with coding agents: several sessions working in parallel on sources, design, the radar and the story, each on its own branch, with a build and a link check deciding what gets published. The source is on GitHub.
To get in touch, see Conversations on the home page. Views here are my own, not my employer's.