Nexus

• 

  Nexus Dashboard

• 

  Nexus Poster

• 

  Product Summary

• 

  Nexus AI Pipeline

• 

  Simulation Panel

Inspiration

A semiconductor plant fire in Taiwan. An export restriction in Brussels. A port strike in Long Beach. Individually, they're headlines. Together, they're market forces — but connecting the dots across geographies, sectors, and supply chains is something only institutional desks with billion-dollar terminals can do.

A 2025 systematic review of 37 papers found that "interactive methods for teaching risk and reward remain limited" and "tools rarely integrate multiple content areas, limiting their ability to convey real-world financial complexity." Only 3 of 37 addressed advanced financial skills. Nexus is a direct answer to this gap (Du, Amor, Ma & Wünsche, 2025 — "Data Visualization for Improving Financial Literacy", ScienceDirect).

What It Does

Nexus is an intelligence platform that maps how global events propagate as financial shockwaves across countries, sectors, and stocks.

It continuously ingests news, classifies events with our own NLP models, and scores every tracked stock using a composite shock formula combining semantic similarity, historical sensitivity, geographic proximity, and supply chain linkage.

1. Interactive 3D globe showing shockwave propagation across countries
2. AI stock predictions with confidence intervals for S&P 100 tickers
3. What-if simulation engine for hypothetical geopolitical events
4. RAG-powered AI financial analyst with live market data
5. Real-time news ingestion and event classification
6. Sector-level impact aggregation and supply chain mapping

How We Built It

• Custom Python model server with sub-100ms inference and no API costs
• BART-large-MNLI for zero-shot event classification
• all-MiniLM-L6-v2 for sentence embeddings and vector similarity search
• 117 per-ticker LSTM models trained from scratch for price prediction with Monte Carlo confidence intervals
• Sphinx NLP orchestration layer with automatic fallback cascading: Model Server → OpenAI → deterministic stubs
• Actian VectorAI DB for vector similarity search and historical event matching
• NestJS backend with 16 modules and Socket.io real-time streaming
• React + Vite + Tailwind frontend with Mapbox GL globe and Recharts
• Deployed on AWS ECS Fargate with CloudFront

Challenges We Ran Into

• Training 117 LSTMs with proper chronological splits required Georgia Tech's Phoenix Supercomputer Clusters
• OpenAI at 200-500ms/article was too slow for real-time — building our own classifier brought it under 100ms
• Staggering WebSocket broadcasts to make simulations feel like real-time events
• Bridging Actian VectorAI DB's gRPC interface to our TypeScript backend via a FastAPI proxy

What We Learned

• Custom models beat API wrappers — sub-100ms inference, for free
• Hand-crafted math (shock formulas, severity curves) is faster and more explainable than expensive LLM calls
• Every external dependency needs a fallback in case of failure (this made our app nearly unbreakable)
• RAG works best with live data, not static knowledge bases. This provides the most up-to-date and informing information

What's Next

• Built to scale — AWS Fargate handles millions of vector comparisons per second
• Satellite imagery, flight tracking, and social sentiment as future data inputs
• Portfolio-level shock scoring and mobile push alerts
• Licensed API for institutional desks
• The geopolitical risk intelligence market is $32B and growing — no one is doing this with AI at this depth
• Information is power, and Nexus is bridging that gap

Built With

• acled
• actian-vectorai-db
• all-minilm-l6-v2
• amazon-web-services
• aws-cdk
• aws-cloudfront
• aws-ecs-fargate
• aws-secrets-manager
• bart-large-mnli
• docker
• fastapi
• finnhub
• gdelt-api
• keras
• lstm
• mapbox-gl-js
• mistral-7b
• nestjs
• newsapi
• node.js
• openai-text-embedding-3-small
• polygon.io
• python
• react
• recharts
• socket.io
• tailwindcss
• tensorflow
• typescript
• vite
• yahoo-finance