Motivation
I created Eule Elli to solve a common problem I encountered during board game nights: constantly pausing to look up rules disrupted the experience. One late game night with lots of reading in the rules I realized that combining retrieval-augmented generation with well-structured game manuals could provide immediate, context-aware answers and dramatically improve gameplay with the emerging abilities of LLMs. The core idea was to replace the static experience of flipping through rulebooks with a dynamic, conversational assistant.
Approach
I designed the project as follows:
- Cross-platform: Built in React Native for web and mobile (iOS and Android).
- Backend powered by Supabase: Using PostgreSQL and PGVector for semantic search of game rules.
- Chunked retrieval: I preprocessed rulebooks by splitting them semantically. This was done by leveraging headings, subheadings, and contextual grouping to enable accurate, context-relevant and cost-effective retrieval.
- Voice-first interaction: Integrated speech-to-text (STT) and text-to-speech (TTS), ensuring players could keep their hands on the game and their eyes on the board.
- RAG pipeline: Initially built on OpenAI’s models and later integrated Gemini Flash 1.5 for better quality and cost savings.
- Document ingestion: Game manuals were consumed as PDFs and processed using OCR (initially using Azure Document Intelligence and later Gemini's Multimodal capabilities).
- Multimodal: I implemented interfaces for text and voice interactions. Users could switch between typing and speaking seamlessly, with real-time feedback to ensure natural communication.
- Security: Authentication and rate limiting were implemented to prevent abuse.
Eule Elli has been accepted in the Microsoft for Startups program.
Publisher Collaboration
I proactively reached out to select board game publishers for explicit permissions to use and process their rulebooks. Negotiating usage rights taught me about licensing, IP compliance, and balancing product feasibility with legal clarity. One publisher was supportive, while another very large publisher was too skeptical to collaborate with for an MVP. They first wanted proof of the usefulness before letting me use their content for an MVP which would have been necessary for public testing. This and the realization that monetization is going to be very hard in the German market with such a product eventually made me stop this project and start working on new things such as WhisperIt.
Learnings
This project deepened my understanding of retrieval-augmented generation pipelines and taught me how to design reliable AI interactions in real-world, conversational scenarios. I also gained experience in:
- Voice-based interfaces and how they differ from text-based UX.
- Handling the trade-offs of chunking strategies in semantic search.
- Balancing rapid prototyping with legal negotiations when working with third-party IP.
Overall, Eule Elli represents how I combine technical depth, user-centric design, and practical business considerations to create AI-powered tools that solve real-world problems.