mein Unterricht

The first AI-Native knowledge graph for teachers

What started as a content migration turned into a 4-year R&D partnership focused on AI-native product development. We helped meinUnterricht consolidate systems, reduce operational complexity, and launch their first contextual AI assistant – all under tight deadlines.
4
Years
of Systematic Experimentation
delivered in
3.5 mth.
production-ready implementation
launched AI assistant with
69.7%
activation rate
Beyond Simple Search
Germany's teaching materials exist in over 150,000 PDFs, Word files & web pages spread across 16 federal states, dozens of publishers & thousands of curricula.
This content was effectively "dark data" – neither machine-readable nor connected.
Together with meinUnterricht (Klett Group), we spent four years on an experimental question: can you build a system that automatically extracts pedagogical structure from unstructured educational documents and connects it into a curriculum-aware knowledge graph – without manual curation?
‍
Standard AI tools couldn't solve this. General-purpose LLMs and search engines are optimized for general text. Applied to educational content, they hallucinate structures, miss domain-specific relations, and produce generic results that ignore the pedagogical context. To bridge this gap, we moved beyond simple search into systematic R&D.

From legacy constraints to AI-Native velocity

Launched the "MUKI" AI assistant with a 69.7% activation rate and over 12% engagement-to-conversion retention in its first month.
R&D-Driven Search Optimization
Developed a proprietary pipeline to transform 150,000+ unstructured documents into a searchable, curriculum-aware Knowledge Graph.
Zero-Disruption Transition
Decommissioned the legacy e-learning platform on schedule, migrating all users, subscriptions, and content to a unified, SCORM-compliant engine.
Delivered production-ready implementation in 3.5 months.
And enabled continuous delivery of new AI features.
Our Approach

4 Years of Systematic Experimentation

03
Designing a pedagogical otology from scratch
We started by solving a fundamental data problem: the lack of a common language.
We defined a complex web of relations, including curriculum alignment and prerequisite knowledge, to map how educational concepts actually interact across different states and grade levels. This wasn't an adaptation but rather the creation of a new, scalable framework capable of disambiguating pedagogical intent across thousands of heterogeneous sources.
The Hybrid Extraction Pipeline
To bridge massive format variance and regional terminology gaps (e.g., Lernziel / Kompetenzerwartung), we built a multi-stage data extraction pipeline.
03
  • Rule-based parsers for document segmentation – identifying structural elements (chapters, sections, task blocks) across wildly different PDF and Word formats from different publishers
  • Entity recognition for pedagogical concepts – detecting learning objectives, competency references, and topic markers in natural language
  • Domain-Specific Fine-Tuning: Since standard transformers (BERT, GERBERT) lacked educational specificity, we built and annotated a custom German educational corpus to fine-tune our models.
03
The Core Uncertainty: Semantic Relation Extraction
Deriving implicit relations via LLMs was highly experimental; it was unknown if prompt engineering alone could resolve the inherent pedagogical ambiguity.
Systematic Experimentation: We tested hundreds of prompt variations and few-shot learning sets on Claude (via AWS Bedrock).
The Challenge: Some approaches captured connections but introduced systematic misclassifications (e.g., inventing non-existent prerequisite chains). We had to find the exact "golden" prompt strategy to ensure the system correctly maps foundational concepts across different competency levels without expensive manual training.
Graph Architecture & Query Optimization
The extracted knowledge required a storage layer that could handle complex, cross-curricular queries (e.g., "Show me all Class 7 materials that are prerequisites for topic X").
We implemented this on PostgreSQL with graph extensions and pgvector. Since educational query patterns don't match standard database workloads, we had to experiment with graph traversal strategies and index optimization to ensure the system performs at scale.
03
Our Features

The AI Assistant "MUKI"

To prove the value of our R&D, we launched an AI assistant directly within the document reader. Unlike generic chatbots, MUKI is curriculum-aware.
Results from the 2,000-user pilot group:
Contextual Precision
Teachers confirmed that the assistant’s ability to "read" PDFs and map them to curricula felt fundamentally different from standard AI tools.
69.7%
Activation Rate
Crushing industry benchmarks for new feature adoption.
12%+
Retention
High engagement-to-conversion rates in the first month.
Contextual Interaction
Teachers can "chat" with any PDF, asking questions based strictly on the document’s content.
Automated Material Prep
Generating quizzes and teaching prompts that are tailored to the specific material and its competency targets.
Zusammenfassung schreiben: Inhalt kurz und verständlich zusammenfassen
Curriculum-Aware Summaries
Summarizing documents while maintaining the pedagogical structure and curricular context.
Generating quizzes and teaching prompts that are tailored to the specific material and its competency targets.
Custom File Context
Users can upload their own files (lesson plans, notes, images) as additional input for the assistant, preserving full AI functionality across personal workflows.

Team perspective

"The hardest part wasn't building the system – it was proving that reliable, automatic extraction of pedagogical relations from unstructured content is possible at all. Every pipeline stage had moments where we weren't sure the approach would work. Four years in, we have a knowledge graph that captures structure no general-purpose AI can see. That's genuinely new."
Kamil Łochyński
AI Engineer at Vazco

Technical Foundation: Why AWS Bedrock?

The research required a robust infrastructure that balanced power with strict constraints:
Data Residency
Bedrock ensures all data remains in Europe, a non-negotiable requirement for the Klett Group and German educational standards.
Rapid Iteration
Allowed us to pivot quickly during the 4-year experimentation phase without managing complex inference servers.

Next steps: Scaling the AI Layer

The embedded assistant is just the beginning. With its foundation validated and user engagement far exceeding benchmarks, we're now building toward a system-wide, AI-native teaching experience. Here's what comes next.
Integrating knowledge-graph-powered features
into search, dashboards, and content recommendations across the platform
Expanding extraction coverage
across additional publishers, content types, and curricular frameworks as model reliability improves
In-line editing & export for AI-generated outputs
– turning quizzes, prompts, and summaries into ready-to-use teaching materials
Voice interaction
for mobile lesson preparation, designed for teachers planning between classes

Let’s explore how we can create a similar impact for your business.

Michał Zacher, CEO at Vazco
Book a free consultation today
ServicesExpertiseOur WorkPodcastsTeam
Vazco P.S.A.
Podwale 7/1
50-043 Wrocław, Poland
Contact
[email protected]
+48 609 257 363
NIP: PL8971878028
KRS: 0000961394
REGON: 385903063
Privacy policy
© 2025 Vazco