Hmm, interesting... kind of what I align with... if the reason you are using an LLM is for 'reasoning', go ahead, but going through this post, i think my best bet would be to use deterministic code for the most part, and perfecting my predictions using an ML model only. Relying on an LLM only is risky, as mentioned. Even if you are going to implement a RAG architecture, The database of that RAG pipeline in itself could be used to create deterministic code and possibly training of the ML model. I guess LLMs must be used only at a stage after the two above-mentioned processes, and only for reasoning purposes and not as a primary decision-maker, maybe human-in-loop?.