~/about
AI builds fast. Someone senior decides what's good.
AI speeds up building enormously. It does not deliver a good product on its own. That takes vision, craft and hard choices, made by someone who has sat on the buyer's side of the table. That is the job here.
$ whoami
Nico Meihuizen
founder · AI-native software engineer
$ history | head -1
age 12 · IBM 5110 · a terminal and a manual
$ cat career.log
2011–2025 enterprise IT sales · IBM · SoftLayer · Oracle · Minitab
2022 Minitab 100% Club
2026 meihuizen.ai · Kaunas
$ ls ./cs-degree
ls: cannot access './cs-degree': No such file or directory
$ ls ./shipped
second-audience/ agent-fritz/ five/ this-site/
~/who you get
One person owns it. Start to finish.
Founder and AI-native software engineer, after fourteen years of enterprise sales. I own the architecture, the process and the hard technical calls, whether the team around it is me and nine agents or your own twenty engineers.
No agency layers, no junior handoffs, no account manager relaying your notes. You talk to the person who writes the code. Having been the account manager for fourteen years is exactly why.
// from an IBM 5110 to enterprise IT
I got my first computer at 12: an IBM 5110. No tutorials, just a terminal and a manual. Coding stayed a hobby through a career selling data center, cloud and SaaS to buyers who could smell a shallow pitch instantly.
That career taught one habit: know the product inside out before you present it. Now I build the product.
~/vision
Your next buyer meets you through an agent.
It reads your site, compares you with three others and answers before anyone picks up the phone. It answers whether or not you gave it anything to answer with. Most companies have not noticed yet. The agent does not care.
Two things follow. Companies have to be readable by the machines that now do the first round of buying. And one senior person with agents has the reach a department used to need. We build for both: sites and data that agents get right, and agents that do real work.
What we sell, we run on ourselves first
~/what we're good at
Four things, done properly.
Agentic products
Multi-agent systems and LLM-driven tools, built with a process disciplined enough that faster also means reliable.
see: Agent Fritz →Automated AI workflows
Agent-driven pipelines that take the repetitive work off your team, with a human sign-off where it matters.
see: Second Audience →Local-first, privacy-first
Software that runs on your own hardware, so your data stays yours. Built in the EU, for the GDPR you actually live with.
see: One Acre, Zero Dependency →Full-stack web platforms
Architecture, design, deployment and upkeep. This site is written for people and for machines, in English and Lithuanian.
see: the builds →~/how we ship
Fast, with the brakes still on.
Speed from AI is only worth something if the process catches what AI gets wrong. So every change runs the same route, whether it is a typo or a new agent. No exceptions for "small" ones.
- branchEvery change on its own branch. Main is never edited directly.
- pull requestDescribed, linked to the why, readable a year later.
- ci: scan + testsCodeQL security scan and unit tests on every push and pull request.
- ci: evalsFritz answers his trap questions on every preview deploy. Red means no merge.
- human reviewNico reads the diff. Nothing merges unread.
- cd: deployMerge to main ships it. No hand-copied files.
- monitorPublic status page. If it breaks, you see it too.
CodeQL · this sitenpm test · agent-fritz-guardrails, every pushFritz evals · every Vercel deployESLint + Next.js build · Fivesite diagnostics · SEO and security sweep
Code, prompts, evals and site copy. If it is not in Git, it does not exist.
A red check on a preview means it does not merge. Nobody overrides it to make a deadline.
A fixed set of trap questions every agent has to pass before a change goes out.
Credentials live in the hosting platform's settings, per environment.
A privacy page that says what is kept, how long, and who to ask to delete it.
Fritz's guardrails and test harness are a public repo. Read the tests yourself.
~/how I work with AI
A tool. A very good one.
- 01It is a tool, not the workA model on its own produces nothing a client can use. The result comes from the workflow around it.
- 02Pipelines beat promptsOne prompt is a gamble. Controlled inputs and repeatable steps are a production method.
- 03Default output is a draftIt ships when it survives review, not when it compiles.
- 04Know where it breaksEvery model fails somewhere. Knowing where decides whether AI is the right tool at all.
- 05The model does not decide what is goodIt widens what one person can build. The judgement stays human.
~/the crew
One human. Nine agents. One dog.
You don't need thirty employees any more. One person with agents has the reach that used to take a department. Pretending otherwise is the old org chart with better marketing.
The one who steers
The one who fetches
The ones who build
- Claudebuilds, writes, and argues back
- ChatGPTsecond opinion, and register A in a scan
- Geminithird read on the same page
- Perplexityretrieval, and where a claim came from
- Kimilong context, and curiosity
- Mistralfifth register, and the European one
- Copilotinline completion in the editor
- BigMLpredictive models, and new here
- Agent Fritzanswers from these pages only, and fetches the leads
Tool names say which tools. No affiliation, partnership or endorsement implied.
~/the mobile lab
The office has wheels.
There is no office to visit, on purpose. The work happens where the work is: at your table, on a call, or in the van parked outside. It runs fully off grid, so a discovery day in your car park and a build sprint in a Lithuanian forest run on the same power, the same uplink and the same Bialetti.
Anywhere in the EU, at your door within 48 hours.
- POWER
- EcoFlow Delta Pro 3, plus a Heko station for DC
- SOLAR
- roof panels, charging while we work
- UPLINK
- Starlink Mini
- BACKUP
- 3.3 kW inverter generator
- DESK
- ThinkPad, 24" and 14" screens, one cable
- HEAT
- a Grizzly wood stove, built for Lithuanian winters
- CREW
- 1 human, 1 dog
- COFFEE
- 3 a day, Bialetti, brewed on the wood stove
- GRILL
- a Cobb smoker, for the client who stays for dinner
- BAR
- full, opens when the build passes CI
~/where the company is
Registered. And the old line stays.
I changed my mind for three reasons. An MB costs almost nothing to hold open. VAT and an entity turned out to be the price of entry to the conversations I want, not the reward for winning them. And Second Audience stopped being an idea.
A page that was never wrong is a page nobody ever edited.
No company until there is a client. An empty company to look established is what I'd flag on your site.
MB Meihuizen AI, a Lithuanian MB, company code 308157412. The VAT number goes here the day it lands.
check the code at Registrų centras →
> this page is also written for machines. Every fact here is in plain text and structured data, so I find out what agents get wrong about me before you do.
~/next
15 minutes with the person
who builds it.
--:-- in Kaunas · office hours · Mon to Fri, 10:00 to 18:00
"The measure of intelligence is the ability to change." Albert Einstein