From Black Belts to Small Language Models: My Taekwondo AI Journey
- Category
- App Dev
- Format
- Session — 40 minutes
- Audience level
- Intermediate
What happens when you combine an interest in Taekwondo with Flutter, local AI, embeddings, and retrieval-augmented generation? In this talk, I’ll share my journey building an AI-powered Taekwondo application designed to answer questions using a specialized body of martial arts knowledge. What started as an experiment with large language models quickly became a much deeper exploration of how to make AI answers accurate, grounded, fast, and practical on a mobile device. We’ll look at the evolution of the application—from early experiments with cloud and local models to building a RAG pipeline using Flutter, Gemma, embeddings, and Supabase. I’ll discuss some of the problems that turned out to be much harder than expected: choosing what context to retrieve, handling small language models, improving retrieval quality, preventing hallucinations, and determining whether a bad answer was caused by the model, the prompt, or the data supplied to it. Along the way, I’ll share the approaches that worked, the ones that definitely didn’t, and the lessons I learned about building domain-specific AI applications outside of a demo environment.
Speakers
Kevin Moore
Kevin Moore is a Google Developer Expert in Flutter, author, international speaker, and longtime mobile application developer. He has spent more than 14 years building Android applications and has worked extensively with Flutter to create applications across mobile, desktop, and web. Kevin is the author of *Mastering Flutter* and has contributed to several other books on Flutter, Android, Kotlin, and Kotlin Multiplatform. He has also written numerous articles and tutorials and regularly speaks at developer conferences and community events. More recently, Kevin has been exploring the intersection of Flutter and artificial intelligence, building AI-powered applications that combine large and small language models, retrieval-augmented generation, embeddings, and domain-specific knowledge. H