Practical AI insights for European SMEs. Real numbers, real tools, real results.
A statistic appeared in my published research logbook. It looked like a measurement. It was a hardcoded string in a print statement — nothing in the code computed it. Here is how it got there, what happened when I re-ran the experiment properly, and why the obvious conclusion was also wrong.
Researchers built a deliberately useless method that beats the standard score everyone reports. I re-ran it on my own CPU in eight seconds and got their number back. Then I found their published code doesn't reproduce their own paper. Here is what that means when a vendor quotes you a benchmark.
A tool I wrote scanned 65 websites in 83 seconds and reported six GDPR violations. Three of them were innocent. Here is why it lied, the four checks that caught it, and the uncomfortable pattern in which cases a cheap scanner gets wrong.
Generic cloud AI keeps leaking Brazilian Portuguese into European-Portuguese copy — and prompting it to stop does not reliably work. So we fine-tuned a small guardrail model to rewrite text into clean PT-PT. Here is the dataset trick, the 12-minute training run on a 2018 GPU, and the honest results.
Most 'AI research agent' demos are a thin wrapper around a frontier cloud API. Ours isn't. Here is the architecture of a local agent that searches, reads and verifies entirely on hardware we own — the runtime, the self-hosted tools, the subagents, and the proof it does real work.
You do not need 24GB of VRAM or a cloud subscription to run a real AI agent. Mine runs a 12B model on an 11GB RTX 2080 Ti — fully offline. Here is the exact memory budget, the serving stack, and the honest limits, step by step.
I asked a self-hosted model for a cited research brief. It looked perfect — and a third of the citations were pages it never opened. Here is the deterministic check that catches it, the one a prompt rule (and my own eyeballs) missed, step by step.
An AI wrote a migration plan for our infrastructure. It was 95% right — and the missing 5% would have broken production. Here is the independent-verification method that caught it, step by step, with the real findings.
45 automations, 91 LinkedIn posts, 638 leads, 23 repositories. The complete output of building a consultancy with an AI coding assistant.
From writing code to reviewing it. From keeping context in my head to documenting everything. Four concrete workflow changes after 400 hours with an AI coding assistant.
One VPS, all open-source tools, zero SaaS subscriptions. The complete self-hosted stack behind ImparLabs.
A double newsletter send, invisible broken tracking, and 14 silent workflow failures. Real incidents from building 45 automations with AI.
Two months, €214 total cost, 19 documented incidents. What really happens when you build an entire consultancy infrastructure with an AI coding assistant.
Weekly AI insights for European SMEs. No hype, just what works.