FinSight AI Agent 🏦

A production-grade financial intelligence agent built with LangGraph + AWS Bedrock (Claude) — designed for regulated fintech environments.

---

What This Does

FinSight is an agentic AI system that handles complex financial analysis queries through a multi-step reasoning graph. It combines:

• Real-time transaction anomaly detection — flags unusual spending patterns
• AML/OFAC compliance screening — checks transactions against watchlists
• Risk scoring — dynamic risk assessment with explainable outputs
• Natural language financial reporting — generates executive-ready summaries
• Guardrails enforcement — blocks PII exposure and hallucinated financial data

Built for Capital One-style regulated environments where data governance, auditability, and compliance are non-negotiable.

---

Architecture

User Query
    │
    ▼
┌─────────────────────────────────────────────────────┐
│                  FastAPI Gateway                    │
│              (auth, rate limiting, logging)         │
└────────────────────┬────────────────────────────────┘
                     │
                     ▼
┌─────────────────────────────────────────────────────┐
│              LangGraph Agent Graph                  │
│                                                     │
│  ┌──────────┐    ┌──────────┐    ┌──────────────┐   │
│  │  Intent  │───▶│  Route   │───▶│   Execute    │   │
│  │  Parser  │    │  Tools   │    │    Tools     │   │
│  └──────────┘    └──────────┘    └──────┬───────┘   │
│                                         │           │
│  ┌──────────────┐    ┌─────────────┐    │           │
│  │  Guardrails  │◀───│  Synthesize │◀───┘           │
│  │   Checker    │    │  Response   │                │
│  └──────┬───────┘    └─────────────┘                │
│         │                                           │
└─────────┼───────────────────────────────────────────┘
          │
          ▼
┌─────────────────────────────────────────────────────┐
│           AWS Bedrock (Claude 3.5 Sonnet)           │
│                                                     │
│  ┌──────────────┐  ┌─────────────┐  ┌───────────┐   │
│  │  Knowledge   │  │  Guardrails │  │   Model   │   │
│  │    Bases     │  │  (PII/harm) │  │Evaluation │   │
│  └──────────────┘  └─────────────┘  └───────────┘   │
└─────────────────────────────────────────────────────┘

Agent Tools

Tool	Purpose	Compliance Relevance
transaction_analyzer	Detect anomalies in spending patterns	Fraud Prevention
risk_scorer	Dynamic multi-factor risk assessment	Credit Risk / BSA
compliance_screener	AML/OFAC watchlist screening	Regulatory (BSA/AML)
report_generator	NL → structured executive reports	Audit Trail
account_profiler	Behavioral baseline + deviation scoring	KYC / CDD

---

Quick Start

Prerequisites

• Python 3.11+
• AWS account with Bedrock access (Claude 3.5 Sonnet enabled)
• AWS CLI configured (aws configure)

1. Clone & Install

git clone https://github.com/yourusername/finsight-agent.git
cd finsight-agent

python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate

pip install -r requirements.txt

2. Configure Environment

cp .env.example .env
# Edit .env with your AWS region and settings

3. Run the API

uvicorn src.api.main:app --reload --port 8000

4. Try It Out

# Analyze a transaction
curl -X POST http://localhost:8000/agent/analyze \
  -H "Content-Type: application/json" \
  -d '{
    "query": "Analyze account ACC-4821 for unusual transactions in the last 30 days and flag any AML concerns",
    "account_id": "ACC-4821"
  }'

Or run the interactive demo:

python examples/demo.py

5. Docker

docker compose up --build

---

Project Structure

finsight-agent/
├── src/
│   ├── agent/
│   │   ├── graph.py       # LangGraph graph definition
│   │   ├── state.py       # Typed agent state (Pydantic)
│   │   ├── nodes.py       # Graph node functions
│   │   └── tools.py       # Agent tool definitions
│   ├── bedrock/
│   │   └── client.py      # AWS Bedrock client wrapper
│   ├── api/
│   │   └── main.py        # FastAPI application
│   └── utils/
│       └── logger.py      # Structured logging
├── tests/
│   ├── test_agent.py
│   └── test_tools.py
├── examples/
│   └── demo.py
├── Dockerfile
├── docker-compose.yml
├── requirements.txt
└── .env.example

---

Key Design Decisions

Why LangGraph over vanilla LangChain? LangGraph gives you explicit control over agent flow — cycles, conditional edges, checkpointing. For regulated environments, you need to know exactly what path the agent took and why. LangGraph's state graph makes every decision auditable.

Why AWS Bedrock over direct Anthropic API? In a Capital One-style environment, customer data can never leave your VPC. Bedrock runs inside your AWS account — no data retention, FedRAMP High compliant, full CloudTrail audit logs. The Guardrails integration also blocks PII before it reaches the model.

Why Pydantic v2 for state? Type-safe agent state catches bugs at development time, not in production. Pydantic's validation also documents the contract between graph nodes — readable as spec, enforced at runtime.

---

Extending This

• Swap mock data for real Snowflake / DynamoDB calls
• Add Bedrock Knowledge Bases for internal policy RAG
• Wire in AWS EventBridge for async agent triggers
• Add LangSmith tracing for production observability

---

License

MIT — use freely, attribute appreciated.