01
RAG architecture
Answers come from sourced material, not from a model's statistical haze. RAG structurally reduces hallucinations because answers are derived from curated sources and reviewed by guardrails.
Platform
Sovereign AI infrastructure for sensitive audiences.
We start with AI for children, because it carries the strictest requirements. What works here carries every other sensitive-audiences vertical. The platform brings together RAG architecture, dedicated servers in DACH data centres, in-house moderation and a multi-guard safety stack. No black box, no third-country data. Every one of our solutions runs on this foundation.
Safe AI. Safe Data. Made in Germany.
The technology
The technology behind the platform.
02
Dedicated servers, DACH hosting
DACH hosting, no US hyperscaler, no third-country transfer. ISO 27001-certified data center operator.
03
In-house moderation & Multiguard
Moderation before the answer, not after it. Multiple independent safety layers against harmful answers, parasocial bonding and data leakage.
04
AI-Act-ready & ISO/IEC 42001
Architecture and processes built “AI-Act-ready by design”. ISO/IEC 42001 certification in preparation. AI-Act-compliant across all stages: GPAI since 02 August 2025, standard obligations from 02 August 2026, health high-risk from 02 August 2027, education high-risk from 02 December 2027 – we are ready.
AI transparency
RAG claim
Generative AI, only in curated spaces.
Generative AI is deployed only inside curated knowledge spaces, defined roles and auditable guardrails. Answers come from sourced material, not from a model's statistical haze. RAG structurally reduces hallucinations because answers are derived from curated sources and reviewed by guardrails. Partner IP stays protected. Brand voice stays consistent.
Safe AI. Safe Data. Our Infrastructure.
Data sovereignty for sensitive audiences.
Sensitive audiences need particular attention. Children, especially under 14, are under heightened legal and societal protection. That is the basis of how we act. We align the platform to it, and every HillcrownAI solution builds on this foundation.
User data sovereignty
Conversations, learning histories and interaction data stay in Germany. No third-party clouds, no model training on children's data, no shadow profiles. Parents receive an overview without surveillance.
Customer data sovereignty
Partner content, IP and configurations remain the partner's property. No model training by HillcrownAI or third parties. At the end of the contract: export file, full platform decommissioning.
Infrastructure data sovereignty
DACH hosting in ISO 27001-certified data centres. In-house moderation, Multiguard stack. No US hyperscaler, no third-country transfer. Open-source foundation models, self-hosted.
Legal basis
- AI Act Art. 5 (1) b
- Prohibition of AI systems that exploit vulnerability based on age.
- AI Act Annex III, Education
- High-risk classification for AI in educational contexts from 2 December 2027.
- JMStV and JuSchG
- Age categories 0/6/12/16/18 and protection from content that may impair development.
On this legal basis the platform architecture is built. On this architecture every HillcrownAI solution builds.
Hardware under our own control.
Our own GPU servers. Our own control.
Models, inference and moderation run on our own GPU clusters in German data centres. Multiple cards per cluster, sized to load and growth. We don't share it with cloud providers and we don't rent compute slots on US platforms.
Inference
Nvidia H200
Current generation for the foundation models, operated as a dedicated inference cluster. Carries user responses in real time.
Multiguards
Nvidia H100
Workhorse cluster for the parallel moderation layers of our Multiguard stack.
Image generation and image analysis
Nvidia Blackwell RTX 6000
Dedicated cluster for image-generation and image-analysis workloads that run alongside the language pipeline.
Three reasons for own hardware
- Data control
- Conversations don't leave the German data centre. No token-forwarding contracts with US hyperscalers.
- Latency
- Inference and moderation run physically close to each other. Real-time responses under two seconds remain plannable.
- Compliance
- ISO 27001-certified operator, GDPR-compliant, AI Act-ready by design. No third-country transfers in the architecture.
Hardware sovereignty is the prerequisite for data sovereignty. Both stay with us, so it can stay with your users and your content.
Research & Development
Research and development.
What an AI for children from age four has to look like is an open research question. From today's general-purpose language models, an age-appropriate protected space cannot be derived, neither in interaction, nor in content moderation, nor in latency under real conditions.
We experimentally investigate how cognitive, social and motor developmental stages have to be reflected in user guidance and AI response behaviour, how real-time responses under two seconds can be achieved without data storage, and how closed model loops can be prevented from carrying biases or misinformation into a protection-worthy user group.
The work takes place in a closed, German-hosted research and development environment, together with developmental psychology, pedagogy and AI safety research. Results are open-ended and validated in iterations against scientific and technical risks.
What the platform means for partners
Whitelabel. API. Custom.
Every solution runs on the platform. Partners can build whitelabel versions, take API licenses or have custom implementations developed via HillcrownAI Labs. One infrastructure, multiple paths.
Compliance disclosure
Compliance is a platform job.
AI Act, ISO/IEC 42001 and ISO/IEC 27001 are not just promises, they are architecture. This disclosure line applies to all solutions because they all run on the same platform.
Safety
Infrastructure
AI-Act-ready (architecture and process) · ISO/IEC 42001 in preparation · Data-centre operator ISO/IEC 27001 certified
AI Act key dates at a glance
- 2 Aug 2025
- Prohibited practices already in force (e.g. manipulative systems targeting children).
- 2 Aug 2026
- General-purpose AI and governance obligations apply broadly.
- 2 Dec 2027
- Specific high-risk obligations for educational AI apply (Annex III Education).
HillcrownAI maps use cases along these dates. For education providers this means: the architecture is already AI Act ready today, but the full high-risk obligation programme for education only kicks in at the end of 2027. We take on the documentation and risk structure on a use-case-specific basis.
One platform for every world of childhood.
We build the safe AI worlds for partners in education, edutainment, culture, sport and family. Reach out – we'll get to the configuration in 30 minutes.