We are an AI research and consultancy company from Frankfurt, Germany, and the team behind c:node.
Our work starts from one question: how do you build AI that has to justify its answers? Public administrations, banks and the German Mittelstand cannot act on "the model said so". Every statement needs a source someone can check, every decision a record, and nothing should leave the system without a person approving it.
| Answers grounded in a continuously growing knowledge graph of EU and German law, public registers and market data — with the norm, register entry or document each statement rests on. | Agents propose, people decide. Mails, tasks and changes in connected systems stay drafts until someone approves them, and every decision lands in an audit trail. | EU hosting, GDPR and the EU AI Act built in, pseudonymisation before any external model call, and an on-premises edition down to air-gapped setups on open-weights models. |
| c:node · platform | The product: a workspace that answers from your own documents and systems plus the graph, shows the source for each statement, and holds anything with an outside effect until a person approves it. Cloud, on-premises or air-gapped. Try it · Graph · an MCP server for other tools to query |
| c:node Graph · research core | The continuously ingested knowledge graph behind the answers — EU and German law down to the article, public registers, procurement, market data. Grown under the rules written up in governed-graph-growth. |
| Zulano · public sector | Applications and funding research for public services. A data mesh of more than 34 million structured records as the grounding layer, and a rule-based service that routes a request to the responsible authority deterministically rather than by a model's guess. |
| AI Playbook · self-service | A trilingual AI readiness check for organisations starting out: a questionnaire in, a report out, with the GDPR and EU AI Act obligations that apply to what they described. |
| Whitelabel instances · pilots | The platform runs as a separate tenant per customer — own branding, own graph, own data boundary — for municipal and enterprise pilots. |
| Consultancy · discovery | Where it starts for most customers: a paid discovery that establishes whether the problem is one evidence-bound AI actually solves, and what it would cost to find out. |
flowchart LR
Q["Question"] --> R{"Retrieve"}
G[("c:node Graph<br/>law · registers · tenders · market")] --> R
D[("Your organisation's<br/>documents & systems")] --> R
R --> A["Answer<br/>with a source per statement"]
A --> H{"Action with<br/>external effect?"}
H -- "no" --> U["Answer to the user"]
H -- "yes" --> P["Proposal card"] --> AP["Human approval"] --> X["Executed + logged"]
Live counts in the observatory at graph.c-node.ai, and the method behind them in governed-graph-growth.
We publish methods, not implementations: how we evaluate, govern and grow systems that answer with evidence, mapped to public standards so others can apply and check them.
| Publication | What it is | Cite |
|---|---|---|
| grounded-ai-evals | A four-layer method for evaluating AI systems that answer with evidence — graph correctness, ontology hygiene, answer groundedness, baseline comparison — plus a release gate that blocks on named checks instead of a weighted mean. | |
| governed-graph-growth | A technical note on keeping a continuously ingested, provenance-first knowledge graph trustworthy as it grows: source authority tiers, identity before enrichment, rule-classed linking, schema re-validation, growth health signals. |
Requirements, acceptance and escalation sit in Germany. Delivery runs through an international, multilingual team, which is how a mid-sized budget buys senior people rather than a smaller number of them.
A ticket is the specification and carries its acceptance criteria. Agents implement with access to the repository, the database and the APIs; people review and test; a confirmation gate stands before anything irreversible. Architecture decisions are written down with the option that was rejected and why — a team's memory is what survives the people who were in the room. Scrum or Kanban, with refinement before planning rather than inside it.
| Role we staff | What that means here |
|---|---|
| Data science and machine learning | M.Sc. level. Modelling, evaluation design, and the discipline of asking what a metric is actually measuring before anyone reports it. Our published evaluation method came out of this work. |
| AI engineering | Agentic systems, knowledge graphs, retrieval that has to cite, document automation, and the grounding and provenance layer underneath all of it. |
| Enterprise architecture | Services with one job each, integration into landscapes that already exist, identity and billing as their own systems, cloud or on-premises down to air-gapped. |
| Technical product ownership | The interface between business and engineering: requirements, prioritisation, architecture decisions, stakeholder and committee work in regulated settings, and the release gate. |
| UX/UI and product design | Research, flows and design systems — the interface people actually have to use, not the one that demos well. |
| Full-stack engineering | React, TypeScript and Node at the front, Python and FastAPI behind it, Java and Spring Boot where the existing landscape requires it. |
| Cloud and operations | AWS, Google Cloud, Hetzner and on-premises with local GPU inference. CI/CD, and the running of it afterwards. |
Teams are put together per engagement, from one specialist alongside your people to a delivery unit that owns a product end to end.
- Pilots — public administrations, financial institutions and mid-sized companies that want AI they can audit: start at c-node.ai, or book thirty minutes.
- Research — universities and research institutes working on grounded AI, legal knowledge graphs or human oversight of AI agents: we are open to joint work and co-authored publications.
- Contact — contact@creativate.tech · Leonardo Bornhäußer, CTO
Trust is good. Evidence is better.

