CFO Copilot
AI-Powered Personal Finance Management Platform
An AI-powered financial advisory platform using multi-agent orchestration to analyze financial documents and generate actionable insights. The system processes natural language queries, executes dynamic Python code for analysis, and delivers HTML infographics and React-compatible Plotly charts through conversational interfaces.
Technical Stack
- Backend: FastAPI with async/await, Cloud SQL PostgreSQL
- AI/ML: OpenAI GPT-4o, LangGraph for multi-agent workflows
- Vector Database: ChromaDB with text-embedding-3-large (1024 dimensions)
- Frontend: React 18 with TypeScript, Vite, Tailwind CSS
- Visualization: Plotly for interactive charts
- Infrastructure: Google Cloud Run, Cloud Storage, Vercel
Multi-Agent Architecture
LangGraph Workflow: Five specialized agents orchestrate query processing:
- Query Analyzer: Intent classification and routing
- Data Retriever: Semantic search across financial documents
- Code Generator: Dynamic Python code creation for analysis
- Code Executor: Sandboxed execution environment
- Response Formatter: Output formatting with visualizations
Smart Caching: Similarity-based response caching with 10-minute TTL and 85% threshold, achieving 65% cache hit rate
Key Features
Natural Language Processing: Conversational interface for financial queries with intent classification and entity extraction
Document Processing: Multi-format ingestion (PDF, Excel, CSV, JSON) with OCR fallback for scanned documents
Dynamic Code Execution: Secure sandboxed Python environment (pandas, numpy, plotly) with AST validation and memory/time limits
Interactive Visualizations: Generates HTML infographics and React-compatible Plotly charts embedded directly in conversational responses
Subscription Management: AI-powered recognition of subscription renewals with lifecycle monitoring and cost analysis
Performance Metrics
- <1 seconds for similarity-based cached queries, <30 seconds for new deep analyses
- 50+ simultaneous users supported with horizontal scaling
- 95%+ code execution and user query success rate