My AI twin runs 100% in your browser. Yours could too.
Local AI in practice: I fine-tuned a small LLM on myself and it runs 100% in the visitor's browser. No cloud, no US provider, nothing leaves the device. How it works, and what running AI locally means for company data.
There's a chat bar at the top of this page. Ask it something. "Why not just use Azure?" works well. What answers you is not ChatGPT behind a curtain. It's a small AI model, fine-tuned on me, running inside your browser tab. You can open your network monitor and watch: after the one-time download, nothing goes out. Pull your network cable mid-conversation and it keeps talking.
Why I built it this way
I build and consult on sovereign, GDPR-compliant AI for companies in the DACH region. Every consultant says "your data stays yours." I wanted my own website to prove it in the most literal way possible: the AI you're talking to physically cannot phone home.
The difference matters more than it sounds. When your teams use a US-hosted AI tool, the question is not whether the provider is trustworthy. It is which law can reach the data. Here is the same conversation, in three setups:
| Public ChatGPT | US cloud, "EU region" | On-device / your infrastructure | |
|---|---|---|---|
| Where your words go | US provider's servers | EU datacenter of a US provider | They don't leave |
| Which law reaches them | US law | US law (CLOUD Act follows the provider, not the server) | Yours |
| Who can be compelled to hand them over | The provider | The provider | Nobody but you |
| Running cost | Per seat, forever | Per seat or per token, forever | You own it |
The middle column surprises people. A US court has ordered AI chat logs preserved, including deleted ones, and Microsoft's own legal director testified in 2025 that EU customer data cannot be guaranteed to stay beyond US reach. An EU region answers where the servers stand, not whose law applies.
How it works, briefly
For the technically curious, the recipe in one paragraph: a small open model fine-tuned on ~12,000 question-answer pairs about my work, quantized and exported to run on the browser's own GPU. Facts live in a tiny retrieval index next to the model, so updating what it knows means editing a 90 KB file, not retraining. The model is public if you want to look under the hood. And because small models can still get things wrong, every answer carries a note pointing to the real me.
What local AI (on-device AI) actually means
Local AI (also called on-device AI or edge AI) means the model itself runs on the hardware in front of the user: a laptop, a phone, a server in your own rack. Nothing is streamed to a provider's API. That is different from "private cloud", different from an "EU region", and very different from a public chatbot. The model on this page is the smallest useful example: it lives in the browser's cache and computes on the browser's own GPU. For companies, the same idea usually lands one step up: a self-hosted model on your own infrastructure, behind your own firewall, under your own law.
Until recently this was impractical, and that is why most companies never considered it. Two things changed. Small open models became genuinely capable when they are fine-tuned for a specific job instead of asked to know everything. And the tooling for running them (on browsers, on modest servers, on existing hardware) matured to the point where a working system takes weeks, not a research project.
The cost question nobody runs the numbers on
Cloud AI is priced like a subscription treadmill: per seat or per token, every month, forever, with the price set by someone else. An owned model inverts that. You pay once to build and fine-tune it, it runs on hardware you already understand, and the marginal cost of the next thousand questions is roughly electricity. For a small team the subscription is fine. For hundreds of employees, or for AI baked into a product you sell, ownership crosses over fast, and it removes a risk that never shows up on the invoice: your provider changing prices, terms, or model behavior under you.
Frequently asked questions
Is a small local LLM as good as ChatGPT?
At everything? No. At the one job it was fine-tuned for? Usually yes, and often more consistent, because it answers from your documents instead of improvising. The twin on this page knows my work; it cannot write your marketing plan. That trade is exactly right for enterprise use cases, where you want a specialist under your control rather than a brilliant generalist under someone else's.
Isn't an EU region of a US cloud provider GDPR-compliant?
Compliance and sovereignty are different questions. An EU region settles where the servers stand. The CLOUD Act obliges US providers to produce data they control regardless of where it is stored, and Microsoft's own legal director confirmed under oath in 2025 that EU customer data cannot be guaranteed to stay beyond US reach. If your risk analysis ends at "the datacenter is in Frankfurt", it ended one question too early.
What hardware does self-hosted AI need?
Less than the hype suggests. This page runs a fine-tuned model on whatever laptop you are holding. Serious document or product-data workloads typically run on a single GPU server, and plenty of real work fits on CPU-only machines you already own. The model size is a design decision, not a given: right-sized beats over-sized on cost, speed and auditability.
What does a project like this cost?
It depends on scope, which is why I don't quote figures on a website. The honest path: a free first consultation, then a lean pilot that produces a working system on your data in weeks, with real numbers attached. You decide with evidence, not with a slide deck.
Can my company get a chatbot like this one?
Yes, and usually better: your version knows your products, your documents and your processes, speaks your languages, and runs where your data already lives. The one on this page is deliberately small so it can live in a browser tab. Yours gets a real server.
What this means for your company
The twin is a demo of a pattern, and the pattern is the actual product:
- Document intelligence that reads your contracts inside your infrastructure, so the contracts never become someone else's training data.
- Product-data AI that enriches and classifies your catalog (PIM, ETIM) on your servers.
- Internal assistants your employees can actually be allowed to use, because the answers to "where does the data go" and "which law applies" are: nowhere, and yours.
If a 211 MB model in a browser tab can do this, imagine what runs on one proper server in your rack. A pilot takes weeks, not quarters: a working system on your data, inside your landscape, and real numbers to decide with.
Try the twin: the chat bar at the top of this page. German works too. Same model, same device, and the network tab stays silent.