The author developed Vicquant, an offline financial advisor app, after a competition inspired a unique approach to financial reasoning. Vicquant's intended design features a dual local-model setup, utilizing Llama-3.1-8B-Instruct for deep analysis and Phi-3.5-mini as a faster fallback, all running locally on the user's device via llama.cpp. This split was necessary because optimizing for both reasoning quality and response speed with a single model is challenging. The architecture and logic for this offline functionality have been built and tested on hardware with 8GB RAM. However, the current hosting plan is insufficient to run these local models in production. To make the app functional, the author connected Vicquant to OpenRouter, enabling it to serve AI responses via the internet for the deployed demo. This means the live demo currently requires an internet connection and offline mode is disabled. The author is transparent about this change to maintain credibility and highlight the common problem of the gap between designed and deployable architecture due to cost. The app's core financial analysis, including indicator calculations, signal generation, and backtesting, remains unaffected as it uses deterministic Python, not AI computation. The backtesting gate, which requires signals to prove themselves against historical data, also remains in place irrespective of the AI model used for explanations. Local inference will be reinstated once the hosting cost constraint is resolved, and an update will be posted at that time.