
Surf-tour outfits, weekend markets, the cafés that empty out one Tuesday and overflow the next — these businesses get maybe a day or two of warning when demand flips. A storm wipes out the surf class. A yoga festival lands across the bay. Airbnbs sell out. The owner decides right then whether to add staff, restock, or message the people already booked. Most are guessing. OpsScout watches for them. Five CrewAI agents (a Forecaster, a Demand Modeler, a Logistics one, a Comms one, an Operations Manager) run on Qwen3-30B-A3B-Instruct via vLLM on an AMD MI300X. They pull from Open-Meteo for weather, PredictHQ for events, and an @openbnb MCP server for Airbnb supply and price. That last one's where the model gets interesting. When listings are scarce and prices have jumped, the tourists are already in town and stuck. Bad weather doesn't send them home; it sends them out looking for something to do indoors. That's not a moment to cut hours. The agents' output runs through a LangGraph state machine with Postgres checkpoints, which condenses everything into one ActionProposal: three concrete actions across staffing, inventory, and comms, with cost estimates and a plain summary. The graph then stops at an await_approval node. Nothing fires until the owner hits Approve in the Streamlit dashboard. When they do, the same checkpoint thread picks up where it left off and runs execute. Solo build for the AMD × lablab.ai hackathon. Targeting Best Overall and Best of Qwen.
10 May 2026