End project: a repo-aware bot built by climbing the decision ladder: prompt first, then RAG over your private Playwright framework (qwen2.5-coder + nomic-embed-text + LanceDB, all via Ollama, all offline), then a LoRA adapter on MLX when you want the model to sound like your team. Press Play to learn when each rung wins.