A2Genesis GmbH
We build software, websites and applications — from the first sketch to the running system. Our core discipline is AI-native architecture: systems designed out of intelligence rather than having it bolted on.
Your solution
used where it is measurably better
Services
Seven services, one standard: what we build keeps running without us. AI goes in where it makes a difference — and stays out where it does not.
Where we stand
Right now every application is getting a language model wired into it so the box can say “AI”. We think that is the wrong reflex. A great many tasks are solved faster, cheaper and more reliably in plain deterministic code — code that does the same thing every time and can be read line by line.
So we decide per task, not per proposal. Where a model is a clear advantage, we build it in and measure the gain. Where a handful of rules will do, we write a handful of rules — and we will tell you so even if you came asking for AI.
We prefer to run models on our own hardware. Not out of hostility to the cloud, but because confidentiality, control over the whole system and the freedom to experiment matter to us. What suits your project best is something we decide together.
The difference is not how much AI is in it, but whether anyone checked where it actually helps.
What we do
We work across the full range and have one core discipline. What a task needs is decided by the task — not by what happens to sell well.
The daily craft: applications, websites, interfaces, automation of recurring steps, plus findability and brand presence. Most of it needs no models at all, and that is as it should be — deterministic code is faster, cheaper and the same every time.
Software in which intelligence is part of the architecture rather than a feature bolted on. The difference is testable: an AI-native system keeps context across steps, makes intermediate states visible, and can be recomposed in its capabilities. A retrofitted system has an input field.
We run models on our own hardware and use it to try out what we do not yet know. Two reasons: confidentiality, because nothing leaves the building, and freedom, because on a system you own you can measure everything and change everything. Whatever moves from that into communication is labelled as a research finding, not as an available capability.
Difference
| Compared to | Our position |
|---|---|
| Intelligence added after the fact | AI-native — Intelligence is embedded, not attached. |
| AI as a selling point | Justified — A model goes in when it is measurably better. Otherwise it does not. |
| Selling tools instead of advising | Open-ended — We will also say no when the effort would not pay off for you. |
| Monolithic approaches | Modular — Capabilities can be combined, replaced and reused. |
| “Magical” systems | Traceable — Origin, state and effect stay visible. |
Have a go
Six everyday tasks. Each time, decide whether a language model is the better choice or ordinary code. Two of them are traps — and those same traps get sold elsewhere as AI projects.
Task 1 of 6
Read the amount and invoice number from 4,000 incoming invoices — all in the same format from the same supplier.
How we work
Whichever service — the order is the same. It is here because the most common unspoken question is not “what does it cost” but “how does this start”.
An hour, no invoice. You describe what needs deciding; we say whether we are the right people for it. The no comes here, not after the quote.
In writing, with an effort range and the open questions beside it. Where we cannot prove something, it says so there — not in a footnote.
The smallest part that changes something goes live. It is the measure for everything after it: what does not hold here will not hold at scale.
In stages, each with a result of its own. After any stage you can stop without being left with a building site.
Source code, credentials, build instructions, operating notes. Whether we then take on operations is a separate decision, not a consequence.
In numbers
Counted in our own files, not estimated. Where each number comes from is stated beneath it.
95
Components in the design system
exports from A2 Lumina
171
Design tokens
colours, spacing and type scales in tokens.css
40
Icons in the set
our own icon set, one SVG
16
Pages in A2 Apex
the tenant platform, in operation
Hands-on
Four places where you can check for yourself instead of taking our word for it. Everything runs in your browser, nothing goes to a server.
Six stations straight through a language model — from splitting a sentence to the question of where the computing actually happens. No prior knowledge, just scroll.
Come alongThe·invoice·arrived·yesterday.Four pieces to work yourself. Type a sentence and see why German costs more than English — and each piece states what is computed and what is merely shown.
Try itTwo pieces with a score: guess how many parts a sentence breaks into, and find the bug in the source. Both compute for real, both can be lost.
PlayOur design tokens, the icon set and the shader behind this site. No sign-up, no address, under MIT.
Take a lookTechnology
Not an allegiance to a vendor, but the tools we actually work with. What is listed here is in our own projects too.
The list is information, not a boundary. Where an existing system calls for something else, we work with that — a tool that fits the house is worth more than one that fits us.
Who we serve
Assistant
The assistant knows exactly what is on this page — nothing more. It answers questions about our position, our fields of work and how to reach us, and it says so when an answer is not supported here.
The assistant is not configured on this installation. info@a2genesis.de
Contact
A short message is enough. You get an honest assessment — usually the same day.