🏦 Alex - AI Financial Planning Platform

Production-grade multi-agent system for intelligent portfolio analysis and retirement planning

---

🎯 Overview

Five AI agents. One orchestrated system. Real financial analysis at scale.

Built on AWS serverless with Terraform, powered by Azure OpenAI GPT-4o, and running vector search at 90% less cost. This isn't a prototype—it's production infrastructure handling multi-tenant workloads with autonomous research, real-time insights, and zero server management.

---

✨ Core Features

🤖 INTELLIGENT MULTI-AGENT ORCHESTRATION

• 5 specialized AI agents collaborating through SQS messaging
• Azure OpenAI GPT-4o powering intelligent analysis
• Parallel Lambda execution for 3x faster processing
• Autonomous research with Tavily API integration
• Each agent focused, independent, unified insights

⚡ PRODUCTION-GRADE INFRASTRUCTURE

• 100% Terraform-managed serverless AWS architecture
• S3 Vectors for 90% cost reduction vs traditional vector DBs
• Aurora Serverless v2, Lambda, App Runner, SageMaker embeddings
• CloudWatch observability + LangFuse agent tracing
• Zero servers, automatic scaling, pay-per-use

🔐 ENTERPRISE-READY FULL-STACK

• Next.js 15 + React 19 with TypeScript and Framer Motion
• Clerk authentication with multi-tenant PostgreSQL isolation
• API Gateway with key-based security
• Real-time streaming responses and CloudFront CDN
• EventBridge automated research scheduling (every 2 hours)

---

🏛️ System Architecture

AWS Services Workflow

Job Lifecycle

AWS Infrastructure

• Compute: AWS Lambda (6 functions), App Runner (Researcher service)
• Database: Aurora Serverless v2 PostgreSQL
• AI/ML: Azure OpenAI GPT-4o, SageMaker Serverless (embeddings)
• Storage: S3 (vectors, static assets), ECR (Docker images)
• Orchestration: SQS, EventBridge Scheduler
• Frontend: CloudFront + S3, API Gateway
• Monitoring: CloudWatch, LangFuse (observability)

Database Schema

Schema Design:

• users: Clerk authentication with retirement goals and target allocations
• accounts: Investment accounts (401k, IRA, taxable) with cash positions
• positions: Holdings per account with quantity and valuation date
• instruments: Shared ETF/stock reference data with JSONB allocation metadata
• jobs: Async analysis tracking with dedicated JSONB fields per agent output (no merging logic needed)

---

🛠️ Tech Stack

Backend & AI

• Language: Python 3.12 with uv package management
• LLM: Azure OpenAI GPT-4o (chat completions API)
• Agents Framework: OpenAI Agents SDK
• Embeddings: SageMaker Serverless (sentence-transformers)
• Web Search: Tavily API
• Observability: LangFuse for tracing

Frontend

• Framework: Next.js 15 (React 19, TypeScript)
• Styling: Tailwind CSS
• Auth: Clerk
• Charts: Recharts
• Animations: Framer Motion

Infrastructure & DevOps

• IaC: Terraform (modular architecture)
• Cloud: AWS (Lambda, Aurora, S3, CloudFront, API Gateway, SQS, EventBridge)
• Containers: Docker, ECR
• Database: PostgreSQL (Aurora Serverless v2)
• Vector Storage: S3 Vectors

---

📁 Project Structure

alex/
├── backend/                 # Python agents and Lambda functions
│   ├── planner/            # Orchestrator agent (coordinates all agents)
│   ├── tagger/             # Instrument classification agent
│   ├── reporter/           # Portfolio analysis and metrics
│   ├── charter/            # Data visualization and charting
│   ├── retirement/         # Retirement projections
│   ├── researcher/         # Autonomous web research (App Runner)
│   ├── ingest/             # Document processing and vector storage
│   ├── database/           # Shared database library
│   └── api/                # FastAPI backend for frontend
│
├── frontend/               # Next.js application
│   ├── pages/             # Route components
│   ├── components/        # React components
│   └── lib/               # API client and utilities
│
├── terraform/             # Infrastructure as Code (modular)
│   ├── 2_sagemaker/      # SageMaker embedding endpoint
│   ├── 3_ingestion/      # S3 Vectors, ingest Lambda
│   ├── 4_researcher/     # App Runner research service
│   ├── 5_database/       # Aurora Serverless v2
│   ├── 6_agents/         # Multi-agent Lambda deployment
│   ├── 7_frontend/       # CloudFront, S3, API Gateway
│   └── 8_enterprise/     # Monitoring and dashboards
│
└── scripts/               # Deployment automation

---

🚀 Deployment Architecture

Infrastructure Provisioning

The entire AWS infrastructure is managed through Terraform modules, deployed in sequence:

1. SageMaker → Serverless embedding endpoint
2. Ingestion → S3 Vectors, document processing Lambda
3. Researcher → App Runner service with Tavily integration
4. Database → Aurora Serverless v2 cluster
5. Agents → Multi-agent Lambda functions + SQS orchestration
6. Frontend → CloudFront distribution, S3 hosting, API Gateway
7. Enterprise → CloudWatch dashboards and monitoring

Multi-Agent System

Planner Agent (Orchestrator):

• Receives user requests via API Gateway
• Coordinates execution across specialized agents
• Manages state and context flow
• Returns unified responses

Specialized Agents:

• Tagger: Classifies financial instruments (stocks, bonds, ETFs)
• Reporter: Analyzes portfolio performance, risk metrics, allocations
• Charter: Generates interactive visualizations and charts
• Retirement: Projects retirement scenarios with Monte Carlo simulations
• Researcher: Conducts autonomous web research using Tavily API

---

📊 What I Built

Technical Achievements

✅ Serverless Multi-Agent System: Deployed coordinated AI agents with sub-2s response times

✅ Cost-Optimized Infrastructure: Reduced vector storage costs from $300/mo to $30/mo

✅ Production-Grade Observability: Integrated LangFuse for complete agent tracing

✅ Infrastructure as Code: 100% Terraform-managed, reproducible deployments

✅ Real-Time Web Research: Automated market research with Tavily API integration

✅ Full-Stack SaaS Platform: Clerk authentication, multi-tenant PostgreSQL, responsive UI

✅ Autonomous Scheduling: EventBridge-triggered research updates every 2 hours

✅ Enterprise Security: API Gateway authentication, user-level data isolation

---

🎓 What I Learned

AWS Serverless at Scale

• Architecting multi-Lambda orchestration with SQS
• Optimizing Aurora Serverless v2 for cost and performance
• Managing serverless cold starts and connection pooling

AI Engineering in Production

• Building reliable multi-agent systems with error handling
• Implementing observability with LangFuse tracing
• Prompt engineering for consistent financial analysis

Infrastructure as Code

• Modular Terraform architecture for complex systems
• Managing state and dependencies across 7 infrastructure layers
• Automated deployment pipelines with uv and Docker

Vector Search & Embeddings

• Implementing S3 Vectors for cost-effective similarity search
• SageMaker Serverless endpoint optimization
• Chunking strategies for financial documents

Full-Stack Development

• Next.js 15 App Router with server components
• Real-time streaming from Lambda via API Gateway
• Clerk authentication with multi-tenant isolation

---

🔮 Future Enhancements

Technical Improvements

• WebSocket Integration: Replace polling with real-time bidirectional communication
• Redis Caching: Cache research results and agent responses
• A/B Testing: Compare agent prompts and model performance
• GraphQL API: More flexible data fetching for frontend

Feature Additions

• Portfolio Optimization: ML-powered asset allocation recommendations
• Risk Scoring: Real-time risk assessment dashboard
• Document Chat: Interactive Q&A with uploaded financial documents
• Mobile App: React Native companion app

Infrastructure Evolution

• Multi-Region Deployment: Global availability with Route53
• Kubernetes Migration: EKS for advanced orchestration
• Prometheus/Grafana: Enhanced monitoring stack
• CI/CD Pipeline: GitHub Actions for automated testing and deployment

---

👤 Author

Harsh Patel
📧 code.by.hp@gmail.com
🔗 GitHub • LinkedIn

---

🎓 Special Thanks

A big thank you to Ed Donner for his excellent MLOps Track, which played a key role in shaping this project and my understanding of building production-grade, agentic AI systems.

📘 Course: Generative and Agentic AI in Production
🔗 https://www.udemy.com/course/generative-and-agentic-ai-in-production/

---

⭐ Star this repo if you found it insightful