sykang16/agentic-wealth-intelligence

1# Agentic Wealth Intelligence System2 3## Project Overview4 5## What We're Building6This project implements a **Multi-Agent Wealth Intelligence System** where specialized AI agents collaborate to provide comprehensive financial advisory services.7Agentic Wealth Intelligence system with 3 modules:81. Asset Management (holistic view + LLM queries)92. Investment Profiling (conversational slot-filling)103. Real-time Recommendations (RAG + MCP)11 12### Problem Statement13- **Current Challenge**: Financial data is fragmented across multiple platforms14- **Gap**: Simple data visualization without actionable insights15- **Solution**: Evolve from passive dashboards to proactive AI-driven financial advisor16 17## Core Business Goals181. **Holistic Asset Management**: Unified view of user's complete financial status and provides LLM-powered natural language interface19- Example20```21User: "What's my current liquidity ratio?"22System: "Your liquidity ratio is 0.35. You have $15,000 in liquid assets 23        (cash + checking) against $43,000 in total current assets. 24        This is below the recommended 0.5 threshold."25```26 272. **Intelligent Profiling**: Natural conversation-based investment tendency assessment and advise based on profiling result and real asset allocation28- Example29```30Agent: "Tell me about your investment goals."31User: "I want to save for retirement but I'm worried about market crashes."32Agent: "I understand you're risk-averse. How far away is your retirement?"33User: "About 30 years."34Agent: "That's helpful! With a long time horizon, you can afford more 35       volatility. On a scale of 1-10, how would you feel seeing your 36       portfolio drop 20% in a year?"37```38 393. **Real-Time Advisory**: Hybrid recommendations combining static knowledge with live market data40- Example41```42User: "Should I invest in tech stocks?"43System: 44- Considers: Your conservative risk profile + 30-year horizon +45             current 60% tech allocation46- Retrieves: Latest FOMC minutes, sector analysis from Vector DB47- Fetches: Current tech sector performance via yfinance (live quotes + news)48- Recommends: "Given your already high tech exposure (60%), consider49              diversifying. Recent Fed signals suggest rising rates,50              which historically pressure tech valuations..."51```52 53## System Architecture54 55**LangGraph Hybrid Supervisor Routing** — the orchestrator uses a two-tier Router (keyword fast-path + LLM fallback) that feeds into a flat outer graph. Recommendation queries are handled by an inner Supervisor subgraph that iteratively gathers portfolio and profile context before calling the engine.56 57### Outer Orchestrator Graph58 59```60                 ┌──────────────────────────────────────────────────┐61     User ──────►│                    Router                        │62                 │  Tier 1: Keyword match  (fast, no LLM call)      │63                 │  Tier 2: LLM fallback   (T=0, max_tokens=20)     │64                 └─────────────────────┬────────────────────────────┘65                                       │ conditional routing66       ┌───────────────────────────────┼────────────────┬──────────────────┐67       │ portfolio_query               │ profiling      │ recommendation   │ general68       ▼                              ▼                ▼                  ▼69┌─────────────┐           ┌──────────────┐   ┌──────────────────┐  ┌──────────┐70│  Portfolio  │           │  Profiling   │   │    Recommend     │  │ General  │71│  Module A   │           │   Module B   │   │  [Subgraph]      │  │  (LLM)   │72│  AssetAgent │           │ slot-filling │   │  Supervisor loop │  │ fallback │73│  net worth  │           │  13 slots    │   │  Portfolio+Prof  │  │          │74│  allocation │           │              │   │  +RAG+live data  │  │          │75└──────┬──────┘           └──────┬───────┘   └────────┬─────────┘  └────┬─────┘76       └─────────────────────────┴───────────────────┴─────────────────┘77                                                       │78                                              ┌────────▼────────┐79                                              │     Respond     │80                                              │ append history  │──► END81                                              └─────────────────┘82```83 84### Recommendation Supervisor Subgraph85 86Embedded as the `recommend` node in the outer graph. The Supervisor iteratively decides which context to gather before calling the engine.87 88```89  outer state ──► Supervisor (LLM decision · guard: steps <= 5)90                       │                  │                 │91               portfolio_fetch    profiling_fetch    recommend synthesis92               AssetAgent         ProfilingAgent     enriched query93               .process()         .get_profile_      + RAG (ChromaDB)94                                  summary()          + live data (MCP95                                                       -> yfinance)96                       └──────────────────┴─────────────────┘97                                   loops back to Supervisor98                                               │99                                          finish (steps >= 5100                                          or context complete)101                                               │102                                        outer Respond node103```