Today I'm launching Zeptix v0.01. It's a platform that lets anyone turn their own knowledge into a real AI chatbot — on their own domain, trained on their content, answering with source citations, hosted in the EU. No coding, no DevOps. This is the project I'm proudest of, and here is the story behind it.

For the last few months, most of my building time went into one thing: Zeptix. Today, June 1, 2026, the first public version — v0.01 — goes live. It's an early, honest first step rather than a polished 1.0, and I want to be upfront about that. But it's also a real, working product that solves a problem I kept running into, and I'm genuinely proud of where it's landed.

This post is the launch story: what Zeptix is, why I built it, the engineering decisions underneath the simple promise, and where it goes from here.

The problem: ChatGPT doesn't know your business

Large language models are extraordinary at language. What they don't know is your world — your products, your docs, your processes, your tone. The common workaround is a "Custom GPT," but that stays locked inside one platform: no own domain, no hosting in Germany, no built-in monetization, and no real product you actually control. You're renting a feature on someone else's surface.

I wanted the opposite. I wanted a chatbot that:

That's the gap Zeptix fills. It turns your content into a standalone chatbot product, on your terms.

What Zeptix does

The core loop is deliberately simple — the live homepage shows it without a mock-up or fake video:

  1. Create an account. Enter an email, pick a slug, choose a tone. About 30 seconds.
  2. Upload your knowledge. Drag in a PDF, a website or a Notion export. "zep" — the AI behind Zeptix — reads along.
  3. Go live. One click, and your bot runs on your-name.zeptix.io.

From there you can drop the bot onto your website in three modes — full layout, minimal chat-only, or a pop-up widget — set it to public, private team or paid, and let visitors top up credits so your bot can earn for you. You can run a language bot for FAQ and support, or a coding bot that understands your stack and shows changes as a compact diff right in the chat. Both can live under one account.

The pitch I keep coming back to: a real product on your domain, not a link on someone else's platform.

The engineering under the simple promise

"Go live in minutes" only works if a lot of complexity is hidden well. Under the hood, Zeptix is a multi-tenant platform with a real stack:

Answers with sources, not vibes

When someone asks a question, the bot doesn't just free-associate. It searches the tenant's own knowledge base via vector similarity, grounds the answer in the retrieved chunks, and surfaces where the answer came from. That source-citation step is what turns "a chatbot" into "a chatbot you can trust for your business."

Smart routing across frontier models

A lot of tools quietly train a cheap in-house model and call it AI. The result is evasive, choppy answers. Zeptix takes the other path: it routes every request to leading AI labs — OpenAI, Google Gemini, Anthropic Claude and Groq — picking the most suitable model per request, with a seamless fallback when limits or load hit. You get the best available answer, not the cheapest one.

Multi-tenant from day one

Every customer gets isolated data, their own branding, their own subdomain, their own billing. Public endpoints are built to never leak tenant-specific data into shared caches. This is the unglamorous plumbing that makes a SaaS actually safe to run — and it's been some of the most careful work in the whole project.

The system you don't see: zeptix.dev

The chatbot is the visible part. The workshop behind it is zeptix.dev, where four tools take a bot from empty to live without code:

Data protection wasn't an afterthought

Because Zeptix is aimed at real businesses — including German and EU customers — the data-protection story had to be honest and documented, not a marketing line. The platform infrastructure runs at Hetzner in Germany. A DPA under Art. 28 GDPR is available on request. Where subprocessors (CDN and AI inference) sit in the US, they're safeguarded via the EU-US Data Privacy Framework and/or standard contractual clauses, and every provider and location is listed openly. Customer content isn't used to train third-party models. Getting this right is slower, but it's the difference between a toy and something a company can actually adopt.

Why "v0.01" and not "1.0"

Calling this v0.01 is intentional. The product works end to end — you can create a bot, train it, embed it and run it today — but I'd rather ship an honest early version and improve it in public than pretend it's finished. There are rough edges, the roadmap is long, and the pricing is still settling (Starter at €29/month early-bird, Pro at €69, Business at €249, all with monthly credits and no minimum term). What matters is that the foundation is real and the core loop holds.

If you want the visual breakdown of the architecture, features and the live chat demo, the Zeptix project page walks through it. And if you build knowledge-heavy products yourself, you might enjoy how clubstream.io tackled a similar "structure the chaos" problem in a different domain.

What comes next

The near-term roadmap is about depth: richer training and source verification, more ready-made knowledge packs, a marketplace of plug-and-play extensions, and eventually zeptix.ai — a personal AI advisor that operates the Builder, Training, Analytics and Visualizer for you. But that's tomorrow. Today is about getting v0.01 into the world.

If any of this resonates — whether you want your own bot or a custom AI product built for you — go take a look at zeptix.io, and thanks for reading the launch story.