Nexus | Model Risk Engine

Login
Model Risk Index
Active Equities Watchlist
Narrative Correlation Matrix
High Risk Stock Exposure
Narrative Risk Ticker
Portfolio Risk Scanner
Narrative Vector Space

Inspiration

In March 2020, portfolios dropped 30% overnight. The news explained everything — but our trading apps just showed red arrows. No context, no warning, no why. We realized: every algorithm reads markets, but none of them read the world. We built the missing layer.

What It Does

NEXUS is a real-time Model Risk Engine that uses AI to watch 5,000+ news sources and autonomously:

Discover narrative directions — clusters like "Fed Rate Hike" or "SVB Bank Run Collapse" with zero human labeling
Compute a Model Risk Index — one number (0–1) showing how far current events diverge from model assumptions
Map narratives to holdings — type a ticker and see which stories are pressuring that position via cosine similarity
Visualize relationships in interactive 3D vector space ## How We Built It
Backend: FastAPI + SSE streaming, scraping 5,000+ RSS feeds every 2 minutes
AI on Modal: all-MiniLM-L6-v2 on T4 GPU for 384-dim embeddings, Qwen 2.5-0.5B for narrative labeling. Custom batching queue coalesces 32 calls into one GPU round-trip (97% overhead reduction)
Vector DB: ChromaDB for semantic search and ticker-to-narrative mapping
Frontend: Bloomberg-inspired terminal UI with Chart.js, Three.js 3D visualization, and dark/light themes ## Challenges
Built a batching queue to cut Modal call overhead by 97%
Designed cosine-similarity thresholds to deduplicate overlapping news into unified narratives
Implemented graceful fallbacks — if Modal is down, heuristic labeling kicks in instantly
Managed cross-platform dev (Mac/Linux/Windows) with dual startup scripts ## What's Next Personal alert thresholds, historical backtesting, brokerage API integrations, and expanded source coverage (filings, social media, earnings transcripts)