ArthVest: Glass-Box Financial Agents

A multi-agent financial platform powered by Gemini that uses a custom Arize MCP to turn black-box AI into a fully observable, self-healing system.

Comment

💡 Inspiration

Financial AI often suffers from a massive trust issue. When AI agents hallucinate stock advice, the multi-agent architecture becomes a dangerous "black box." We wanted to shatter this box! We built ArthVest to be a fully transparent, self-healing financial analyst that you can actually trust.

🚀 What it does

ArthVest acts as your personal hedge fund analyst. Using a Consensus Debate, specialized sub-agents (Risk, Technical, Fundamental) actively battle out the bull and bear cases for a stock.

The system calculates a final, mathematical conviction score using weighted sub-agent sentiment: $$ \text{Conviction Score} = \sum_{i=1}^{n} (w_i \cdot s_i) $$ (where (w_i) is the reliability weight of sub-agent (i), and (s_i) is their predicted sentiment)

Instead of hiding the math and logic, ArthVest uses a custom Arize Phoenix MCP to turn the entire system into a "glass box." You can inspect the exact prompts and tool calls. If an agent hallucinates, our autonomous Improver Agent diagnoses the trace and automatically patches the prompt—reducing hallucinations by up to 90%! 📉

🛠️ How we built it

  • 🧠 The Brain: LangGraph orchestrates the multi-agent debate, powered by Google Gemini (via Vertex AI).
  • 🔍 The Glass Box: Arize Phoenix combined with a custom MCP Server captures and evaluates the LLM execution traces.
  • ⚙️ The Engine: FastAPI backend deployed on Google Cloud Run with Supabase (PostgreSQL) for persistence.
  • 💻 The Interface: A stunning React / TypeScript dashboard hosted on Vercel.
  • 📊 The Data: Live market data from Finnhub, Alpha Vantage, and FRED.

🚧 Challenges we ran into

  • Debugging the Black Box: Agents occasionally hallucinated financial metrics. Debugging was impossible until we integrated Arize Phoenix to visualize the exact execution tree. 🌳
  • Auto-Healing: Building an AI that rewrites its own core instructions without breaking working prompts was incredibly tricky. 🧩
  • Networking: We faced IP-blocking when scraping market news from Google Cloud Run, which we solved by dynamically falling back to the Finnhub API. 🛡️

🏆 Accomplishments that we're proud of

  • Creating an Improver Agent that auto-patches prompts based on failed traces. 🤖
  • Reducing agent hallucinations by 90%, making the AI financially accountable.
  • Bridging LangGraph and Arize Phoenix via a custom MCP for perfect observability. 🌉

🧠 What we learned

Observability is the absolute foundation of trust in AI. We learned that treating LLM prompts as dynamic, trackable code via Arize Phoenix is the only way to scale multi-agent architectures safely. 📈

🔮 What's next

  • 💸 Live Paper Trading: Integration with Alpaca to autonomously execute high-conviction picks.
  • 🎭 Custom Risk Personas: Adjustable risk appetites (conservative vs. aggressive) for the agent debates.
  • 🛡️ Advanced MCP: Embedding complex financial evaluation templates directly into our Arize Phoenix gateway.

Built With

  • arize-phoenix
  • fastapi
  • google-cloud-run
  • google-gemini
  • langgraph
  • model-context-protocol
  • python
  • react
  • vertex
Share this project:

Updates