💡 Inspiration

Modern global supply chains are brittle, non-linear networks. A drought in Central America restricting Panama Canal drafts, maritime missile strikes near the Bab el-Mandeb chokepoint, or water rationing in Taiwan can trigger factory halts and freight inflation thousands of miles away.

Supply-chain and procurement analysts are continuously inundated with fragmented, unstructured information — news alerts, port advisories, and geopolitical updates. Much of this is noise or redundant commentary, and human teams cannot cross-reference every event against historical precedents and complex supplier graphs in real time.

We built NexusFlow to transform raw, chaotic global disruption signals into structured, evidence-grounded supply-chain intelligence. By pairing real-time vector retrieval (RAG) with NVIDIA Nemotron hosted on Nebius Token Factory, NexusFlow uncovers how a disruption propagates from chokepoints down to specific commodities, enterprise consumers, and industries.


⚙️ What it does

NexusFlow is a full-stack, real-time AI supply chain intelligence platform:

  1. Continuous Ingestion & Deduplication: Listens to global news (NewsAPI) and maritime/logistics RSS feeds, filtering out consumer noise with a two-tier semantic relevance engine and SHA-256 deduplication.
  2. Entity & Relationship Extraction: Uses spaCy Named Entity Recognition (NER) to map locations, organizations, critical commodities, and disruption causes.
  3. Historical Precedent Matching (Vector RAG): Queries a ChromaDB vector database using 384-dimensional semantic embeddings to retrieve historical parallels (e.g., comparing the current Red Sea crisis with the 2021 Ever Given blockage).
  4. Causal Knowledge Graph: Maintains a persistent NetworkX MultiDiGraph that accumulates real-time relationships, tracks node weights, and calculates dynamic time-decayed geographic hotspots across four calibrated severity tiers (Severe, High, Medium, Low).
  5. AI Propagation Reasoning: Invokes NVIDIA Nemotron-3 Super 120B on Nebius Token Factory to synthesize structured impact assessments, identifying 1st- and 2nd-order downstream propagation chains, mechanism overlaps, and actionable mitigations.
  6. Executive Intelligence Dashboard: Displays insights via an enterprise visual interface with an interactive 3D WebGL Night Globe, a 2D vector disruption map with pan/zoom navigation, and a 3-tier supply chain causal flow network.

🛠️ How we built it

  • LLM Reasoning & Intelligence: Hosted NVIDIA Nemotron (nvidia/nemotron-3-super-120b-a12b) on Nebius Token Factory, providing sub-second token latency and strict structured JSON outputs for supply-chain causality analysis.
  • Vector Store & RAG: ChromaDB with an optimized ONNX Runtime executing all-MiniLM-L6-v2 (384-dimensional embeddings) for cosine similarity search over historical disruptions.
  • Entity Extraction & Graph Modeling: Local spaCy pipeline (en_core_web_sm) coupled with a NetworkX MultiDiGraph to accumulate entity interaction weights.
  • Backend: FastAPI, Python 3.11, and Uvicorn running asynchronous pipeline tasks.
  • Frontend: React 19, Vite, and Tailwind CSS designed with a dignified enterprise aesthetic (warm #895129 panels on #EDE8D0 canvas) and Three.js / react-globe.gl for 3D geospatial visualization.

🚧 Challenges we faced

  1. Semantic Noise vs. Real Disruption: Most news headlines mention economic keywords without describing physical supply chain bottlenecks. We engineered a dual-tier relevance scoring system combining heuristic chokepoint rules with low-latency LLM classification to filter out consumer market noise before vector indexing.
  2. Memory Constraints & Cloud Deployment: Standard sentence-transformer libraries pull over 1.5 GB of PyTorch and CUDA dependencies, easily exceeding cloud memory limits. We migrated the embedding layer to ChromaDB's native C++ ONNX Runtime, cutting idle RAM usage by 65% (down to ~250 MB) while accelerating CPU embedding speeds by 2.5× with zero loss in mathematical vector precision.
  3. Harmonizing 3D Geospatial Visualization: Mapping dynamic, time-decayed disruption intensities onto both a 3D WebGL sphere and a 2D vector map required calibrating coordinate projections and ensuring color-coded severity rings and 3D point markers matched an enterprise design palette.

🏆 Accomplishments that we're proud of

  • End-to-End Live Workflow: Building a fully functional pipeline that connects raw RSS/news ingestion $\rightarrow$ NER extraction $\rightarrow$ ChromaDB Vector RAG $\rightarrow$ live Nebius Nemotron inference $\rightarrow$ interactive React visualization.
  • Grounded AI Outputs: Prompts engineered to distinguish hard source evidence from plausible downstream inferences, ensuring analysts get defensible intelligence rather than unsupported hallucinations.
  • Enterprise-Grade UI: Moving away from generic dark/neon sci-fi dashboard tropes in favor of an executive, warm intelligence aesthetic.

📚 What we learned

  • How to harness Nebius Token Factory for ultra-fast, hosted inference of large-scale open models like NVIDIA Nemotron 120B.
  • The power of combining structured graph algorithms (NetworkX) with unstructured semantic vector search (ChromaDB) to ground LLM reasoning.
  • Techniques for streamlining production AI deployments by swapping monolithic frameworks for dedicated ONNX inference runtimes.

🔮 What's next for NexusFlow

  • Multi-Modal Satellite & AIS Ingestion: Integrating live ship-tracking telemetry (Automatic Identification System) to correlate news reports with real-time tanker congestion.
  • ERP & Procurement Integrations: Connecting NexusFlow to enterprise inventory systems (SAP, Oracle) to automatically notify procurement teams when an active component’s raw materials are at risk.
  • Predictive Reroute Simulations: Simulating maritime fuel costs, carbon taxes, and transit day variances when rerouting around geopolitical chokepoints.

Built With

  • chromadb
  • fastapi
  • nebius
  • nebius-token-factory
  • networkx
  • nvidia
  • nvidia-nemotron
  • onnx
  • python
  • rag
  • react
  • render
  • rest-api
  • spacy
  • sqlite
  • tailwind-css
  • three.js
  • vector-database
  • vercel
  • vite
Share this project:

Updates

Submission history