Beyond Black-Box Prompts: Building an Explainable Multi-Agent Financial Intelligence Platform with FastAPI & React

Quantitative finance and retail investing are increasingly turning to Large Language Models (LLMs) as a sort of all-knowing oracle for stock recommendations. However, relying on a single LLM can be problematic. Such models often show confirmation bias, insert financial figures without verification, and lack proper risk management. In the world of actual investment committees, there isn't a single go-to analyst. Instead, there are specialized, competing experts.

In response to these issues, I created a platform named Vibe-Finance. Vibe-Finance is a comprehensive, end-to-end multi-agent financial intelligence and explainability system. The platform is built using a Python/FastAPI backend, a vector Retrieval Augmented Generation (RAG) over corporate filings, and a modern React/Vite/Tailwind client.

The system architecture separates data ingestion, analysis, adversarial cross-examination, and synthesis into distinct layers. This design prevents putting all data through one large prompt. Instead, it breaks down the process into dedicated layers for easier handling and understanding.

The platform features three main agents:

1. The Quantitative Agent, which uses hard mathematical indicators like Wilder's RSI-14, MACD, and EMA/SMA to analyze stock movements.

2. The Fundamental RAG Agent, which parses regulatory corporate filings like SEC 10-Ks, 10-Qs, and earnings conference transcripts using vector embeddings. It ensures zero cross-contamination between unrelated equities and provides citations for users to verify claims.

3. The Behavioral Risk Manager, which safeguards the investor against risk-taking behaviors. It assesses the user's risk profile (Conservative, Moderate, or Aggressive), maximum historical drawdown caps, high-beta concentration limits, and provides pre-trade what-if nudges if an order threatens portfolio balance.

Critical to the system is the Adversarial Debate Engine, which handles conflict resolution when different agents present opposing views. For instance, if the Quant Agent sees a bullish breakout signal but the Fundamental RAG Agent detects margin compression in recent filings, the Debate Engine detects agent friction and runs a structured adversarial exchange to calibrate the overall decision confidence.

Finally, the platform provides an interactive TradingView client for visualization and telemetry. It uses TradingView Lightweight Charts for technical analysis and displays model latency, regulatory chunk counts, and token consumption for transparency. The platform also undergoes strict testing with an automated test suite verifying 21 critical scenarios.