Not a slide deck, not a sandbox demo. The same AI MVP development pattern that took a leading member network from an unreliable LLM prototype to a production retrieval-augmented chatbot serving 1,000+ HumHub members.
A leading member network needed an AI knowledge chatbot grounded in their internal knowledge base. Off-the-shelf OpenAI Assistants with file search had been unreliable and accuracy was critical. The chatbot had to live inside HumHub – the social network their members already use – and be operated by a non-technical team. We delivered Version 1 in four short sprints: system architecture, RAG implementation with semantic search, HumHub integration, full documentation. Total estimated effort 10 working days. Stack: n8n in Berlin for workflow, Qdrant in the EU for vector search, GPT-4 via Microsoft EU sovereignty as the LLM. Optional fully open-source EU alternatives: Mistral Medium 3 as the model, Milvus as the vector database. Version 1 is now in production serving more than 1,000 members, time-aware. The same pattern shaped our work for GDV (AI Knowledge Assistant over tens of thousands of policy documents for 400+ insurance companies), a leading donation platform (AI email agent classifying enquiries and drafting replies with mandatory human review, on N8N and Azure OpenAI, currently in pilot), a leading German association (AI Member Platform combining chat-based discovery with category filters), and Tannenhof Berlin-Brandenburg (Civic Coding-funded AI transcription pilot for therapy sessions on EU-hosted infrastructure, output formatted for German Pension Insurance reporting).
Working software, fast
EU-compliant by default
Built to be maintained
A working AI system
the prototype is the foundation of the production system, not a throwaway demo
Time to value
first version typically delivered in four short sprints, the way a leading member network launched
Maintained by your team
low-code architecture and documentation so non-developers can operate and extend it
Human-in-the-loop
humans review consequential outputs by default, the way the donation platform's AI email agent drafts replies for staff to approve
EU AI Act ready
risk-tiered, GDPR-aligned, with audit trails and citations built into the architecture
1. Architecture and scope
2. Build and iterate
3. Hand over and operate
We bring a decade of impact-tech experience and more than 160 AI projects since 2019. Through our free AI for Impact course, more than 100,000 people have learned how to use AI for the common good. We do not run inspiration days. We run scoping sessions and build engagements that ship, the way we have delivered AI for the organisations below:
A four-sprint prototype is not a guarantee against scope creep. We are explicit about what is in and out at the start of each sprint. If the use case turns out to need more than four sprints to deliver value, we will tell you, not pad the project.
Fine-tuning is rarely the right answer. For almost every prototype we build, retrieval-augmented generation (RAG) against your own documents is the correct pattern. If you have been told you need fine-tuning, ask why.
Production AI is not magic. Every system we build needs monitoring, citations, audit trails and a human in the loop where the cost of a wrong answer is high. We bake these in from the first sprint, not bolt them on at the end.
How long does AI prototype development take?
What is the difference between an AI prototype and an AI MVP?
Should we use fine-tuning or RAG?
Which LLM should we use?
Where does our data go?
Can we keep the system after delivery?
Do you handle regulated data?
Tell us about the use case. We will reply with a proposed scope and a sprint plan, usually within a working day.
Simon Stegemann
Co-Founder and CEO
