Arthvest — AI Investment Research on Aurora & Vercel

Arthvest — AI Investment Research on Aurora & Vercel

Comment

Inspiration

Retail investors increasingly rely on AI tools for stock analysis — but every existing solution has the same fatal flaw: no audit trail. When an AI says "BUY," you can't trace why it made that call. What data did it analyze? What were the disagreements? What validation did it pass? The decision is a black box. In finance, trust is everything — and without a fully reconstructable decision history, trust is impossible.

We set out to build an AI investment platform where every single decision is fully reconstructable from the database.

What it does

Arthvest is an AI-powered investment research platform where Amazon Aurora PostgreSQL is the single source of truth for every decision ever made.

  • Dashboard — Real-time global market indices, market breadth, and categorized news feed
  • Discovery — AI evaluates the broader market with scores for Economic Context, Market Pulse, News Sentiment, and Macro Context. Filter AI-picked stocks by Short, Medium, and Long term horizons
  • Analyse — Run deep, customized analysis on any stock with our multi-agent AI pipeline
  • History — Complete ledger of every AI recommendation stored in Aurora PostgreSQL
  • Audit Trail — The crown jewel: a timeline view that reconstructs the full decision lineage from Aurora — Run → Agent Signals → Multi-Agent Debate → Validator → Final Recommendation

Every recommendation flows through 4 specialist AI agents, an adversarial debate engine, and a 3-layer validator. Every step is persisted in Aurora PostgreSQL as a permanent, queryable audit trail.

How we built it

  • Frontend: Built with Vercel v0 for rapid prototyping, React + Vite + TypeScript + Tailwind CSS, deployed on Vercel
  • Backend: FastAPI (Python 3.12) running on AWS EC2, orchestrating the multi-agent pipeline with LangGraph
  • Database: Amazon Aurora PostgreSQL Serverless v2 — 19 normalized tables with foreign keys, JSONB columns, check constraints, and performance indexes storing 706+ recommendations and 2,800+ agent logs
  • LLM: Amazon Bedrock (Claude 3.5 Sonnet & Haiku) for multi-agent reasoning
  • Audit Trail: A single Aurora JOIN query reconstructs the complete decision chain — which agents voted, what signals they produced, how the debate resolved, and the paper-trade P&L

Challenges we ran into

  • Aurora networking: Configuring VPC, security groups, and SSL connections between EC2 and Aurora was the biggest time sink. Getting sslmode=require working with SQLAlchemy connection pooling required careful debugging
  • WAIT discipline: Building an AI system that confidently says "do nothing" was harder than building one that always trades. 310 out of 706 recommendations are WAIT — the agent exercises restraint
  • Demo pacing: Cramming a full-featured platform into a <3 min video while keeping the database story front-and-center required multiple script iterations

What we learned

  • Aurora PostgreSQL's JSONB support is incredibly powerful for storing semi-structured AI outputs while maintaining relational integrity
  • The audit trail pattern (full decision lineage stored in a relational DB) is applicable far beyond finance — any AI system making consequential decisions needs this
  • Vercel v0 dramatically accelerated UI prototyping for data-dense interfaces

What's next for Arthvest

  • Real-time Aurora read replicas for live portfolio dashboards
  • DynamoDB integration for high-frequency market tick data alongside Aurora for analytical queries
  • Expanding the agent pipeline with sector-specific specialist agents

Built With

Share this project:

Updates