What I’ve Learned About LLMs While Building My First Prototype

I’ve been building the Python/FastAPI backend prototype for a Spanish language-learning game. The pipeline connects Groq speech-to-text, GPT-OSS 120B for NPC dialogue, and Azure AI Speech for spoken responses, with HTTP/REST handling communication with Unity.

My biggest lessons have been structuring prompts with NPC and mission context, keeping API credentials in environment variables, and separating service integrations into modules. Testing each component individually has helped me understand how to turn separate AI services into a working conversation pipeline.