Inspiration
Construction sites run on chaos: WhatsApp threads, buried PDFs, missing submittals, and gut feelings. When materials get ordered, the real problem begins. We realized that answering a simple question like "Where is the structural steel?" requires a project manager to cross-reference five different, disconnected dashboards.
Standard RAG (Retrieval-Augmented Generation) is dead for this kind of problem. If you ask standard RAG about a delayed shipment, it just retrieves the shipping document. It is blind to cascading logic. We were inspired to build a system that doesn't just search for documents—it understands the physical reality of the site. We needed to give the supply chain a brain.
What it does
NexusChain AI is a conversational control tower powered by GraphRAG. It actively ingests scattered, messy data and maps it into a living knowledge graph.
- The Blast Radius Predictor: If a supplier emails about a 10-day fabrication delay, NexusChain automatically updates the graph and alerts the manager. It predicts the exact cascading impact—showing exactly how a delayed HVAC unit shifts Phase 3 framing, blocks the electrical subcontractor, and triggers a crane extension penalty.
- Semantic Traceability: Ask the AI, "Are the lobby fixtures on site?" NexusChain traces the specific lights back to the original RFI approval, checks the shipping node, and delivers a definitive status.
- The Trust Layer: Construction managers do not trust black-box AI. Every time NexusChain provides a generative answer, the UI visually displays the exact graph path the AI traversed. No complex jargon, just deterministic proof of how the AI reached its conclusion.
How we built it
We bypassed standard vector databases and built a dynamic relational architecture.
- Ingestion Engine: We used Gemini to process unstructured data (emails, PDFs, schedules). We constrained the LLM with strict JSON schemas to extract precise nodes and edges without hallucinating.
- Knowledge Graph: We used Neo4j to map the ecosystem. Nodes (Materials, Suppliers, Documents, Tasks) were connected via hard edges (
FABRICATED_BY,DEPENDS_ON,BLOCKS). - Routing Logic: To calculate multi-hop delays, we implemented Mixture of Recursion (MoR). Instead of a single query, the system recursively walks the graph to assess n-tier consequences until it reaches the terminal project milestone.
- Generative UI: We built a frontend that instantly renders the traversed subgraph alongside the text response to form the Trust Layer.
Challenges we ran into
Turning chaotic job site data into clean, structured graph nodes is brutally difficult. Zero-shot LLM extraction failed early on. We had to build a deterministic entity resolution layer before writing to the graph—ensuring that "Apex Metals," "Apex," and "Apex Metls" mapped to one single permanent UUID instead of duplicating nodes and breaking the logic.
Accomplishments that we're proud of
We successfully built a system that actively connects the dots instead of just regurgitating text. We proved that GraphRAG is the ultimate unlock for physical supply chains, and by building the Trust Layer, we solved the explainability problem that kills most AI startups in the construction space.
What we learned
Standard RAG cannot handle the interconnected web of a physical supply chain. We also learned that end-users don't care about the underlying AI models; they care about verifiable truth. Showing the visual node path next to the answer is the strongest way to build instant product credibility.
What's next for NexusChain
Integrating live unstructured external variables. We plan to ingest real-time geospatial and weather data, mapping severe weather alerts directly to the specific transit routes of our material nodes to predict supply chain disruptions before the supplier even knows they are delayed.
Log in or sign up for Devpost to join the conversation.