Answers about our AI applications and working with us.
We develop applications that make company knowledge accessible or automate tasks, and integrate AI capabilities through MCP interfaces. Projects we have implemented include a knowledge system, a document management system, an ERP system with AI capabilities, a graph creation application and a presentation moderator. We describe these project examples without naming customers.
AI app development →Both are possible, depending on the task and existing systems. Our focus is on new applications built with Svelte. MCP interfaces allow internal and external AI assistants to access the provided capabilities; the interface defines which actions are available.
MCP integration →We develop interactively with AI within an established technical framework. UI components in Storybook, reusable packages, a clear project structure and working instructions such as AGENTS.md guide development. Human code reviews and automated tests check the results.
Governed Vibe Coding →Depending on the project, we connect existing databases, files on a file system, chats, emails and dedicated knowledge bases. Users ask questions in a web application or through an AI assistant connected via MCP; the system can show the sources used. It can also summarise information and compare multiple sources. Creating a dedicated knowledge base from this information is a planned extension.
Knowledge integration →We define access to suit the project: individually per user or through a shared account with defined permissions. Our MCP interfaces generally use bearer tokens linked to permissions. In our requirements application, we have implemented individual user permissions with Keycloak and a separate token for each user.
MCP and permissions →Our implemented DMS can index a network drive, process Word, Excel, PDF and PowerPoint files and images, and create summaries, detailed content and embeddings for search. It supports questions such as “What was the amount of the last invoice?” and suggests how to file documents and group related scanned pages. Users review these suggestions; fully automatic processing is a possible extension. A knowledge graph connecting relationships across documents is planned.
Documents and knowledge →We recommend user approval for changes and agree the rules for each task. In our requirements application, AI only creates proposals that users must confirm; AI cannot delete anything, and an audit log records the activity. Simple, low-risk tasks can run automatically, such as file summaries that we have implemented using local AI in a project. Agentic workflows can coordinate multiple specialised agents.
Workflows and agents →We develop for the customer, who decides whether the solution uses local AI or a cloud service. We do not operate AI ourselves or provide specialised data protection advice, but we work with the company’s data protection officers. We define the required capabilities together and document them, for example in our AI-enabled requirements application. For workflows, we use tools including LangChain and LangGraph, with LangSmith to monitor token usage.
Technical implementation →We usually join once the overall direction is established. If it is still open, we offer discovery workshops, organise the results afterwards and work with the customer to define concrete use cases, supported by the connected AI. We then usually develop an MVP that the customer can already use. For novel ideas, a prototype can be a useful intermediate step.
From use case to app →The customer receives the complete source code, documentation and a handover. Server maintenance is the customer’s responsibility; we can continue to maintain and develop the application. Our approach to AI-assisted development also enables customers to make changes themselves.
Development and further improvements →