Portfolio Project — Production-ready codebase demonstrating AI system architecture.
FinLens extracts actionable insights from unstructured financial documents. Analysts spend hours manually reviewing dense SEC filings, annual reports, and regulatory documents—often 1000+ pages of financial statements, risk disclosures, and regulatory language.
FinLens processes these documents through GPU-accelerated parsing and multi-agent AI analysis, delivering accurate answers to natural language queries. What takes hours of careful reading and cross-referencing is accomplished with high accuracy through advanced AI analysis.
FinLens is a portfolio project implementing production-grade AI infrastructure for financial document analysis. The system uses multi-agent orchestration, GPU-accelerated document processing, and scalable cloud services. Code quality and architecture patterns follow enterprise standards, suitable for production deployment.
| Feature | Description |
|---|---|
| 🤖 Multi-Agent RAG | 3-agent LangGraph orchestration for complex financial queries |
| ⚡ GPU Processing | MinerU-powered document parsing on dedicated GPU infrastructure |
| 🔍 Hybrid Search | Semantic + keyword search across vectorized financial documents |
| 💬 Chat | Conversational interface with memory management |
| 📊 Financial Analytics | SEC filing analysis, company comparisons, trend insights |
| 🔒 Enterprise Security | Rate limiting, authentication, data isolation |
| 📈 Scalable Architecture | Railway deployment with Supabase + Qdrant + Redis |
The multi-agent system handles complex analytical queries through specialized workflow stages:
- Query Decomposition: Agent 1 breaks down multi-part questions into sub-queries, retrieves relevant context from conversation memory, and performs hybrid search across documents
- Multi-Document Analysis: Agent 2 analyzes retrieved chunks from multiple documents, extracts financial metrics, and performs cross-company comparisons
- Synthesis & Validation: Agent 3 generates structured responses with citations, creates data visualizations (Chart.js), and validates answer accuracy
Supported Query Types:
- Cross-company comparisons (e.g., "Compare Tesla's R&D spending vs Ford and GM")
- Multi-year trend analysis (e.g., "Show Microsoft's operating income for the last 5 years")
- Complex aggregations (e.g., "How has cloud revenue growth compared between Amazon, Microsoft, and Google over the past 3 years?")
- Risk factor extraction across document sections
- Financial statement data extraction with calculations
graph TB
A[Next.js Frontend] --> B[FastAPI Backend]
B --> C[LangGraph Orchestrator]
C --> D[Agent 1: Retrieval]
C --> E[Agent 2: Analysis]
C --> F[Agent 3: Generation]
D --> G[Qdrant Vector DB]
D --> H[Supabase Storage]
B --> I[VastAI GPU Server]
I --> J[MinerU OCR Engine]
I --> K[Document Chunking]
K --> L[Embedding Generation]
B --> M[Stripe Payments]
B --> N[Redis Queue]
style A fill:#e1f5fe
style B fill:#f3e5f5
style I fill:#e8f5e8
- FastAPI - High-performance async web framework
- LangGraph - Multi-agent workflow orchestration
- Qdrant - Vector database for semantic search
- Supabase - PostgreSQL with subscriptions
- Redis - Task queuing and caching
- Voyage AI - State-of-the-art embedding generation
- GPT-4o/GPT-4o-mini - Task-based model selection (GPT-4o for analysis/generation, GPT-4o-mini for query processing)
- Voyage-large-2 - Financial document embeddings (2048 dimensions)
- MinerU - GPU-accelerated PDF OCR parsing
- Hybrid Search - Semantic + keyword retrieval with TOP_K_INITIAL=30, TOP_K_FINAL=8
- Next.js 14 - React framework with App Router
- TypeScript - Type-safe development
- Railway - Cloud deployment platform
- Vercel - Frontend hosting with edge functions
- Python 3.12+
- Node.js 18+
- Railway account (for deployment)
- Supabase project
- Qdrant Cloud instance
-
Clone the repository
git clone https://github.com/StephaneWamba/finlens.git cd finlens -
Backend Setup
cd backend python -m venv venv source venv/bin/activate # On Windows: venv\Scripts\activate pip install -r ../requirements.txt cp .env.example .env # Configure environment variables uvicorn backend.api.main:app --reload
-
Frontend Setup
cd frontend npm install npm run dev -
GPU Processor Setup (Optional)
cd gpu-processor pip install -r requirements.txt python main.py
See deployment documentation for Railway + Vercel setup.
- Architecture Overview - System design and data flows
- API Reference - REST endpoints and examples
- Deployment Guide - Production setup instructions
- SEC Filing Analysis: Extract insights from 10-K, 10-Q reports
- Company Comparisons: Cross-company financial metric analysis
- Trend Identification: Historical performance pattern recognition
- Risk Assessment: Financial health indicators and red flags
- Large PDF Parsing: GPU-accelerated processing of 1000+ page documents via MinerU OCR
- Structured Data Extraction: Tables, charts, and financial statements from parsed content
- File Format Support: PDF, PNG, JPG, JPEG (images processed via OCR)
- Batch Processing: Asynchronous document queue with parallel GPU processing
- Semantic Chunking: Intelligent chunking with page/heading boundaries (max 2000 chars)
Contributions welcome. Open an issue or submit a pull request.
This project is licensed under the MIT License - see the LICENSE file for details.
- LangChain - Multi-agent orchestration framework
- VastAI - GPU infrastructure for document processing
- Supabase - Open-source Firebase alternative
- Railway - Modern cloud deployment platform