Aurum — AI Financial Intelligence Platform

Inspiration

Retail investors in India spend hours manually analyzing stock data across scattered platforms. I wanted to build a tool that lets anyone query financial datasets in plain English — no Excel, no finance degree required.

What I learned

• How to build a production-grade RAG pipeline with pgvector and OpenAI embeddings
• How to stream LLM responses token-by-token using Next.js 15 App Router
• How to design composite-indexed PostgreSQL schemas for fast queries on large datasets
• How to secure APIs with JWT authentication and rate limiting at scale

How I built it

AI Layer: RAG pipeline using OpenAI embeddings stored in PostgreSQL pgvector, enabling natural-language queries over financial datasets with token-by-token streaming — reducing perceived latency by ~60%.

Backend: FastAPI with a Moving Average Crossover algorithm for stock trend analysis. Composite-indexed PostgreSQL schema via Prisma cut query time on 100K+ row datasets by ~40%.

Frontend: Next.js 15 App Router with React Server Components for fast, streaming UI.

DevOps: Dockerized full stack with GitHub Actions CI/CD. JWT auth + rate limiting (100 req/min) securing all endpoints.

Challenges

• Optimizing pgvector similarity search across large financial datasets without latency spikes
• Streaming partial LLM responses while simultaneously running ML analysis in the background
• Designing a schema that performs well on both time-series queries and vector similarity search

Built With

• ai
• docker
• fastapi
• github-actions
• jwt
• next.js
• openai-api
• pgvector
• postgresql
• prisma
• python
• typescript
• vercel