In this project-style tutorial, I set up DeepSeek R1 locally using Ollama and put it head-to-head with Google Gemini Pro 1.5 in a practical GenAI system-design workflow. The goal isn’t “which model is smarter” in the abstract—it’s how they behave inside an app: latency, response quality, and how reasoning changes when you add retrieval. I walk through what DeepSeek R1 is (chain-of-thought style reasoning + the GRPO training angle) and what Gemini Pro 1.5 is (Mixture-of-Experts), then I wire both into the same Streamlit interface so the comparison is apples-to-apples. From there, I build two modes: non-RAG chat and a RAG pipeline using FAISS as the vector DB, so we can see how context precision improves when the model is grounded on retrieved chunks. I also test both models with classic logic prompts (Socrates / syllogisms) and more “real” prompts to evaluate reasoning and consistency. The key takeaway: model choice matters, but system design matters more—RAG, prompt framing, and evaluation setup can swing results a lot. If you want a reproducible baseline, I’ve linked the full source code so you can run the same experiment and extend it with your own docs and metrics.