Can users just ask, and trust the answer?
Chat is the interface, not the system. We build assistants connected to real company data and workflows: for customers, internal teams and products in their own right.
Chat is a surface, not a product
In modern LLM and agent architectures, holding a conversation falls out of the design almost for free.
What comes out varies more than the interface suggests. The question isn’t whether you want a chatbot, but what sits behind it and where its answers come from.
Answers based on your data
Existing internal databases. Knowledge bases built for the purpose. Public specialist databases. And the operational systems you already run.
The model's own knowledge and web research are the last resort, never the first. That ordering decides whether your people trust the assistant or quietly stop using it.
Probabilistic or deterministic
A language model is probabilistic: it returns something plausible, not something certain. In a regulated domain, that isn’t enough.
So we build hybrid systems. The model decides what is relevant, deterministic code decides what is true, and where a result can be checked the check overrules the model.
Our clients hire us to build what gives them an edge. Confidentiality is part of the job. Publishing the details could give that edge away. We’ve delivered in regulated environments and on projects worth hundreds of millions.
What we've built
A client product connected to multiple data sources, including legacy systems never built for AI. Built for hundreds of thousands of users in a regulated domain.
Behind one interface sit a knowledge graph, databases, documents, user history and memory. The client’s own experts can steer its knowledge and behavior without a developer.
The classic case, built properly: answers drawn from your actual content and data, not a decision tree that annoys people into calling anyway.
Product knowledge on demand, and a practice partner for the conversations that are hard to rehearse with a colleague.
A budgeting tool for specialists who know more than the model does, built on their own cost data and domain knowledge. That’s the only way it survives contact with people who spot a wrong number instantly.
What makes them hold up
Four questions decide whether a chatbot survives its first month in front of real users.
- Where did the answer come from?An assistant that can’t show its source is a rumor with a good interface.
- What happens when it’s wrong?If the customer notices first, the system isn’t finished.
- Who’s allowed to change what it says?Your own team or an agency on a maintenance contract, and that decides your operating cost for the next five years.
- What does it cost at scale?The number that matters is ten thousand users, not ten.
What we could build for you
- We build assistants that answer your customers from your real data, so your team keeps the cases that need judgment.
- We make what your company already knows findable in one place, with the source attached to every answer.
- We take the routine questions off your experts, so their hours go to the ones only they can answer.
- We give your software a way to be talked to, by keyboard or by voice, so a report becomes a question instead of a dashboard.
- And whatever you have in mind that AI and solid software engineering could make possible.
Under the hood
Most engagements start small: talking it through, a few weeks of engineering, something real you can judge. If it works, we take it further.If it doesn’t, you find out early
Document AI
PDF to product
View themeData Pipelines
Built to scale
View themeAI Content
Quality at volume
View themeN° 06 Contact
Tell me what you’re planning
No form, no ticket system. Email me directly and you get me, not an inbox someone else manages.
Moritz Miedler · Founder, replies personally