When a chip plant shuts down in Taiwan or a battery factory halts in Europe, carmakers face immediate disruption. Supply chain teams end up buried under hundreds of news tabs trying to figure out which factory line will stop first. The news tells you what happened, but it doesn't know your internal enterprise operations like purchase order PO10482 carrying €2.4M worth of chips to Stuttgart with only 10.9 days of inventory left. Asking standard LLMs to do that math risks hallucinated figures, and in manufacturing, a calculation error stops an assembly line. We built CrisisSentinel on a straightforward division of labor: models read and reason; the database does the counting.

What It Does- CrisisSentinel is an autonomous, dual-agent supply chain command center: Live Web Intelligence: Ingests breaking disruption headlines and conducts multi-query research using Tavily. Claim Extraction & Conflict Resolution: Uses NVIDIA Nemotron-3-Nano on Nebius Token Factory to extract verified claims, impacted countries, and component categories. Conflicting reports trigger targeted follow-up searches. Deterministic ERP Exposure: Evaluates financial and operational risk with zero LLM guesswork. A pure SQLite engine joins claims to open purchase orders, calculating total capital at risk , shortest inventory runway across plants, and approved alternative vendors. Strategist Action Plan: NVIDIA Nemotron-3-Super (120B) structures up to three priority reroute actions, strictly constrained to open POs and ERP-approved alternative suppliers. Adversarial Critic Audit: A second independent Nemotron-3-Super agent red-teams the plan checking rule consistency against database figures and flagging operational risks (such as 90% single-supplier volume concentration). Pure SQL What-If Simulator: Operators drag disruption duration and capacity sliders to model projected stockout curves and line stoppage days in real time. Human-in-the-Loop Execution: Reroutes remain strictly proposed until approved by an operator, committing updates directly to the database. How We Built ItOrchestration & Reasoning: Deployed on Nebius Token Factory using tiered NVIDIA models: Nemotron-3-Nano (30B) for high-throughput extraction and contradiction judging, and Nemotron-3-Super (120B) for executive mitigation strategy and adversarial auditing. Real-Time Web Intelligence: Integrated Tavily Search API with a 3-query triangulation strategy and dynamic conflict follow-ups. Deterministic Core: SQLite ERP engine (voltauto.db) modeling a German EV manufacturer (VoltAuto GmbH: 5 assembly plants, 21 tier-1 suppliers, 199 open purchase orders), built with FastAPI and Python 3.13. Frontend Dashboard: React 18, Vite, and Tailwind CSS with Server-Sent Events (SSE) streaming live pipeline execution telemetry.

Challenges We Ran Into- Model Output Strictness: 120B reasoning models in JSON mode can occasionally wrap responses in conversational framing or markdown blocks. We implemented a fault-tolerant parser with retries to enforce strict Pydantic schemas. Dynamic Endpoint Resolution: Nebius Token Factory model IDs have inconsistent casing across tiers (nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B vs nvidia/nemotron-3-super-120b-a12b). We implemented dynamic runtime discovery via models.list(). Eliminating Hallucinated Arithmetic: Grounding mitigation plans required immutable Pydantic bindings where material names, plant IDs, and order values are read-only copies from database records.

Accomplishments That We're Proud Of- 100% ERP Arithmetic Precision: Verified 20/20 checks against an independent SQL oracle with zero discrepancies. High-Accuracy Information Extraction: Achieved 9/9 category extraction (100%) and 8/9 country extraction (88.9%) against live search queries. Rigorous Verification: Built an automated test suite with 35 unit and property tests covering edge cases, failure states, and transactional database integrity.Zero-Latency Fallbacks: Integrated a recorded mode caching real runs so the entire command center functions offline and demonstrates instantaneously.

What We Learned- Complex enterprise multi-agent architectures work best with strict functional separation: keep non-deterministic tasks (summarization, contradiction analysis, challenge formulation) inside large reasoning models, and anchor operational truth and financial exposure strictly within SQL balance-sheet engines.

What's Next for CrisisSentinel - Direct connectors to SAP S/4HANA and Oracle SCM via enterprise webhooks. Automated vendor API/EDI querying to check real-time alternate factory capacity before recommending reroutes. Sub-regional port and logistics corridor geocoding.