SkillDuck (i40) – Improving a Multi-Agent E-Learning Pipeline

Starting point

With SkillDuck, i40 has built an AI authoring tool that generates complete, didactically structured e-learning courses from documents and general knowledge. i40 won the eLearning Award 2025 in the AI category with it.

The existing AI pipeline had two concrete weaknesses:

• Hallucinations in the generated course content
• Inaccurate time estimates in the generated courses

Both cost trust and rework, especially in a tool that is meant to save authors time.

Solution

In a three-week sprint, TAISC analyzed the existing pipeline and improved it in a targeted way. The pipeline works with specialized AI agents for outline, learning objectives, narrative and didactic content. A RAG layer supplies context from uploaded PDFs, and human-in-the-loop cycles ensure quality at critical points.

The goal of the rework: an automatic 80% solution. The AI builds most of the course, authors review and refine the rest.

How to measure the answer quality of such systems instead of relying on spot checks is covered in our post on ground truth data in RAG systems.

Technology stack

• Azure AI Foundry, OpenAI GPT-4o (LLM orchestration)
• Python, LangChain (backend / agent framework)
• RAG architecture with vector indexing (context-sensitive generation from PDFs)
• Next.js, ReactJS (frontend / authoring interface)
• Human-in-the-loop workflows (quality assurance / hallucination reduction)

Result & Impact

• 3-week sprint to analyze and improve the existing pipeline
• Goal: an automatic 80% solution per course instead of drafts that need complete rework
• Hallucinations and inaccurate time estimates identified as concrete weak spots and addressed
• Course creation in hours to days instead of weeks to months is SkillDuck’s promise
• Scalable to any document base and language