I built KaiwaBuddy, a local AI Japanese conversation partner for a friend learning beginner Japanese. The core problem KaiwaBuddy addresses is the difficulty of applying learned vocabulary and grammar in real conversations. It offers English meanings, beginner-friendly grammar explanations, and corrections for common mistakes. The application also provides follow-up questions to simulate natural dialogue. Currently, it focuses on N5 Japanese and Minna no Nihongo Lesson 1 concepts. KaiwaBuddy utilizes a hybrid approach, combining a deterministic learning engine for known patterns with Gemma 3 1B as a fallback AI model. This hybrid system ensures reliable instruction for core concepts while allowing for open-ended interaction. The project emphasizes open innovation by using the locally runnable Gemma 3 1B model via Ollama, avoiding reliance on proprietary cloud APIs. Local inference is crucial for privacy and developer freedom in customizing AI experiences. Working with an open model also demystified the AI system for the developer. KaiwaBuddy is being submitted for the Hacktoberfest Weekend Challenge, highlighting its local inference capabilities. The most significant learning was that reliable AI applications often combine deterministic logic with AI models for structured feedback.