The "Hallucination" Tax in Fintech In the race to build the biggest LLM, we’ve overlooked a critical flaw: Generative AI is a probability engine, not a calculation engine. For most users, a chatbot is simply a shortcut to data. When an investor asks, "What is my 1-year expected return?" , a +/- 2% "hallucination" isn't a minor quirk—it's a financial liability. This is why we built RIIA (Risk Informed Investment Approach) using a deterministic, local-first architecture. Read more about the project here We traded generative creativity for mathematical certainty. The Architecture: Semantic Routing Instead of sending raw text to a massive model in the cloud, RIIA uses a three-layer local pipeline : The Brain (Sentence Transformers): We use all-MiniLM-L6-v2 to map user queries to one of 20 predefined "Investment Intents." By setting a confidence threshold (0.42), we ensure the system only answers when it is certain of the user's goal....