FinTracker AI- Intelligent Finance Router

FinTracker AI is an intelligent financial decision engine built on top of a fully deployed personal finance platform. Instead of routing every query to a single expensive cloud model, we built a classification and routing layer that dispatches each financial query to the most appropriate local model based on task complexity. A spending summary routes to Phi-3 Mini. Trend analysis routes to Gemma 3 4B or Qwen 2.5 7B. Complex forecasting and what-if simulations route to LLaMA 3 8B. Every routing decision is explained in the UI — which model was selected, why, and what task type was detected. The system runs entirely on local inference via Ollama, with Groq as a cloud fallback. Users can also manually override the model selector to choose between Phi-3 Mini, Gemma 3, LLaMA 3, and Qwen 2.5 — all running locally. Before any LLM call, a financial context engine pre-processes the user's real transaction data, account balances, spending categories, and monthly summaries into a structured object. This keeps prompts short, token costs low, and response quality high. The system also computes a four-dimensional financial health score — savings, cash flow, spending stability, and anomaly risk — displayed in real time on the dashboard. The platform is built with FastAPI, MySQL, and session-based authentication. The AI layer integrates directly into the existing codebase, reusing the same DAO layer, auth system, and API structure. No new database connections. The AI endpoints are authenticated, stateful, and work with real user financial data. Streaming responses via Server-Sent Events give users a fast, responsive experience even with larger local models. It is an inference routing system applied to a real financial product, demonstrating that local open-source models can replace cloud LLMs for structured domain-specific tasks.