fix: route LLM to Groq (deepseek-r1) instead of hardcoded gpt-4o/openai

Initial commit