← Back to services

Proven in production: AI search for SURF

Together with SURF and Tim Fröhlich, funded by the GÉANT GN5-2 Incubator, we built hybrid search for orchestrator-core: semantic, fuzzy and structured search on plain PostgreSQL, with a type-safe agent layer on top so the language model never writes raw SQL. It shipped as open source in orchestrator-core v5.0, and every organisation that upgrades runs it.

We use the same building blocks in our own products. Specialist agents for ShopVirge and the Workflow Orchestrator sit behind one chat interface, reach the backends through MCP and carry the user's own login on every call, so the backends check what that person may do.

Read the SURF case

How we build an AI system

Most of our projects share the same shape: agents with a clear job, grounded in your own data through hybrid search, reaching your systems through MCP, on the model that fits the task. It is the setup we run ourselves.

Who uses it
Your teamChat in LibreChat or Slack
Your applicationsOpenAI-compatible API
Other agentsA2A
Your agents · Python
  • A specialist agent per domain
  • Hybrid search & RAG
  • Tools through MCP
  • User identity on every call
  • Evals
Pydantic AI any model
PostgreSQL + pgvectorEmbeddings, text search and filters in one database
What it talks to
Your systemsWFO, ShopVirge or your own APIs via MCP
Your documents & dataIndexed for retrieval
Hosted modelsAnthropic, OpenAI and others
Self-hosted modelsOpen models on your own hardware
Based on our own agent setup

What we do in AI

Search with vector databases

Implement semantic search using state-of-the-art vector databases. Find meaning, not just keywords, enabling smarter retrieval across documents, products, and knowledge bases.

RAG pipeline design & implementation

Design and build Retrieval-Augmented Generation pipelines that ground LLM responses in your data. Reduce hallucinations and deliver accurate, context-aware answers.

AI integration into existing applications

Seamlessly embed AI capabilities into your current stack, from intelligent APIs to smart automation, without rebuilding from scratch.

Production-ready solutions

We focus on practical, scalable AI that works in the real world: proper evaluation, monitoring, and iteration, not just demos.

Frequently asked questions

What's the difference between RAG and fine-tuning, and which do we need?

For most enterprise use-cases where you need answers grounded in your own documents, RAG (Retrieval-Augmented Generation) is the right starting point: it keeps your data outside the model and stays cheap to update. Fine-tuning only pays off when you need a specific tone or format at scale. We help you choose based on your data and budget, not hype.

How do you stop an LLM from hallucinating on our data?

We ground responses in your content using retrieval pipelines and citations, so the model answers from real sources instead of guessing. It's the same principle behind the hybrid search system we built for SURF's Orchestrator-Core.

Do you build on hosted models like OpenAI and Anthropic, or self-hosted ones?

Both. We integrate the leading hosted models when they're the best fit, and deploy self-hosted, open models when data sovereignty or cost matters. It's the same approach behind our OpenClaw work.

Can you add AI to our existing application without a rebuild?

Yes. We embed AI into your current stack through APIs and targeted automation, rather than starting from scratch.

What makes an AI system production-ready rather than a demo?

Proper evaluation, monitoring, and iteration, plus cost control and security. We focus on systems that keep working after launch, not one-off prototypes.

Ready to build something great?

Tell us what you're working on. We'll think along and tell you honestly where we can help.

Get in touch