-
AI in Practice: Four Lessons Slides Won't Show You
Four hard lessons from operating AI: explain a decision, notice silent data, define the numbers, and keep a blocked action blocked.
-
When should an autonomous agent ask again
It was re-approved because the hash changed, not because the risk changed. The line that came out of it.
-
European registers are opening. Just not the ones I measured
The new rules target beneficial ownership registers, not company registers. My near-failing grade survives.
-
Cloudflare changes the AI bot defaults. Block training and you lose Google too
New defaults from 15 September. Blocking training takes Googlebot with it, and you only notice from impressions.
-
From AI Agents to an AI Organization: Who Thinks, Who Acts, and Who Is Accountable?
A layered AI organization separates planning, ML evidence, real-world action, and human accountability.
-
“We’re clean” is a claim. Here is the procedure that turns it into a fact
On 4 August an npm worm hit more than 400 packages in the keyv/cacheable family. Most write-ups report the threat. This one is the seven concrete checks without which “we’re clean” means nothing — plus the two times I broke my own freeze.
-
x402: the first real coin. And four hooks only a live run finds
Last time, only probes arrived. This time a real coin settled on Base mainnet — actual USDC, end-to-end. Honestly: a controlled demo, not a paying customer. Plus four hooks that pass every unit test and only surface against a real counterparty.
-
MCP dropped state. What that means if you run the servers
On 28 July 2026 MCP shipped its biggest spec revision since launch: a move to a stateless core. I run eight public MCP servers — here is what the change actually means for operations, not a recycled press release.
-
I deleted my AI's self-check — and the output got better
I gave my AI more instructions on how to be careful — and made its output worse. The new model generation verifies itself; explicit verification instructions cause over-verification. Why I deleted those parts of my manual, and the one line I kept on purpose.
-
Six models, one brief: what a panel of AIs revealed about my site and themselves
I ran six different AI models on the same subjective task — reviewing my own site's positioning. Each got identical inputs and never saw the others' answers. The verdict mattered less than where they all agreed, where they split, and what one caught that the rest missed.
-
Director and worker: shipping a release one card at a time
I didn't ship a new version of my own product with a single agent. A strong model planned and verified; a weaker one executed exactly one bounded card at a time and returned evidence. A log of the real run — including the two hours one agent returned nothing.
-
Regulators want a defensible "why" from AML agents. Good architecture gives it for free
Regtech 2026 demands explainability from AML/KYC agents — a defensible reasoning chain behind every decision, not just a score from a black box. Build the agent as a set of tools and that audit trail falls out of the architecture. From running due-diligence tools of exactly this kind.
-
How I stopped filling in my mileage log by hand
I gave a language model access to my own mileage-log app. The hard part was not teaching it to parse “Tuesday, Prague to Brno and back”. The hard part was deciding what it must never be allowed to do — and why that boundary is the whole product.
-
x402 is booming. All that reached my servers were probes
x402 has processed over 169 million payments and pulled in Stripe, AWS and Google. My MCP servers have seen none of it — only probes checking whether I am payable. An honest look at what that does and does not prove.
-
Three kinds of AI bot: one of them is your customer
From 15 September 2026 Cloudflare changes its defaults for AI bots. The headlines say “AI crawlers get blocked”. But if you run MCP servers, blocking the Agent category cuts off your own distribution — that bot is not a pest, it is an end user.
-
The AI Act delay didn't cover your chatbot
Article 50 of the AI Act applies from 2 August 2026 and the Digital Omnibus delay did not touch it. If you run a chatbot or an AI agent, you have three weeks. What I'm checking on my own products — including one question I can't answer: who declares the machine when an MCP server never talks to a human?
-
A report card for Europe's business registries: from Estonia's bulk download to Germany's PDF-for-money
How many EU countries can you cover with due-diligence data from free registries alone, without paying for a single excerpt? I wired them into my own tool and handed out grades. Estonia earns an A+, Germany a D-minus — and the least accessible data belongs to exactly the countries people ask about most.
-
An operating manual from an AI model: what survives when the model gets pricey
A post made the rounds: have the expensive model write down how it thinks, then load that into a cheaper one. I took the advice literally and ran the trap test it promised — the hype didn't hold, but something more valuable did: portable discipline, not the model's raw capacity.
-
Local RAG sizing: rent the GPU before you buy it
A GPU worth thousands can be test-driven before you buy it for the price of lunch. A benchmark of local RAG over confidential documents on rented RTX 3090 and 5090 cards from vast.ai — and why half the result had nothing to do with hardware, but with two env vars and one silent default.
-
LLMs as advisors, not oracles
Why I run a panel of AI models to prepare a decision — three independent reads, the disagreements that matter most, and why the decision itself always stays human.
-
Who Pays for the Agent? x402, Cards, and Banks in the Fight for Agentic Payments
The internet never got a native payment layer for machines. In 2025 three camps are fighting for it at once — x402, card tokens, and banks. What each can do, where each fails, and why the decisive layer is identity and mandate, not money transfer.
-
How I Built a Payment QR in One Day — and What Claude Desktop Taught Me
In one afternoon I built a free European payment QR generator — entirely in the browser. What was easy, what surprised me — and a principle that applies to every AI tool: the model proposes, the deterministic layer verifies and bears responsibility.
-
ARES, ISIR, ČNB: We Have the Data, We're Missing the Interface
The Czech Republic has surprisingly good open data. The problem is the interface: SOAP from 2005, zero sandbox, breaking changes without notice. What we can learn from Estonia and what we can't — and why not to wait for it.
-
Six months building an open-source MCP server for Czech compliance data
ARES, ISIR, sanctions screening, and a due-diligence aggregator as MCP servers. Technical gotchas, architectural decisions, and why I added a SaaS layer.
-
How an AI Agent Learned to Read the Czech Business Register
The story of building cz-agents — an open-core ecosystem of MCP servers connecting Claude/ChatGPT to Czech data (ARES, ISIR, sanctions). How the first organic user found the project through the official MCP Registry, and why I chose open-core over full open source.
-
Solo Architecture: Your First Custom LLM Is Almost Certainly a Mistake
Building your own RAG pipeline or finetuning an open-source model sounds like a strategic edge. It's almost always expensive procrastination. The default is paying for an API. This piece is about the one case in ten when it isn't — and the four gates that keep you in the default until you genuinely run into them.
-
The Economics of Dirty Data: When to Deploy Vision AI and When It's a Trap
Classic ETL over unstructured data is brittle and expensive to maintain. Vision AI promises relief — at what cost? A decision framework on four axes (cost, accuracy, scalability, audit), applied to a case study from the Czech government-data domain.
-
The Large Context Window Trap: Why 1M Tokens Are Often a More Expensive, Dumber Solution Than RAG
Stuffing the whole book into a prompt looks like the holy grail. Reality: high latency, astronomical cost, and the "lost in the middle" problem. When is large context the right call and when is it lazy architecture that makes the product worse?