FinSight AI: Coupling a Large Language Model with a Live NoSQL Database for Conversational Financial Analytics and Rule-Assisted Fraud Screening

Financial dashboards are good at showing data and poor at explaining it. This paper describes FinSight AI, an open-source web dashboard that pairs the LLaMA 3.3 70B model, served through the Groq inference API, with a live MongoDB Atlas transaction store so that an analyst can question the data in plain English and receive answers grounded in the actual database contents. On every request, the backend aggregates the user's transactions into a compact statistical context — monthly revenue and expense series, category totals, the largest credits and debits, and portfolio-level risk — and injects it into the model's system prompt, which keeps responses tied to real values rather than model guesses. A lightweight rule engine scores each incoming transaction from 0 to 100 using amount and recipient heuristics and flags entries at or above a threshold of 40; the language model then supplies natural-language explanations of the flags on demand. The prototype, built on React 18, Node.js/Express, and JWT plus Google OAuth 2.0 authentication, loads its five dashboard endpoints in under 800 ms and returns AI analyses in 1.5–3 s on a free-tier database. We report the design, the prompt-construction method, observations from a seeded 60-transaction dataset, and an honest account of the prototype's limitations. The full source code is public under the MIT license.