Inspiration
We were inspired by how quickly a single automotive supply-chain disruption can affect an entire network. A delayed shipment can change production schedules, delay dealer deliveries, and leave customers waiting. We wanted to connect those consequences and turn incident reports into clear, explainable response plans.
What it does
Our prototype uses linked supply-chain data to trace supplier delays, plant outages, and quality holds to affected operations. A coordinator works with three specialists focused on supply, dealers, and customer care and quality. Together, they produce dry-run plans for rescheduling production, transferring inventory, notifying dealers, prioritizing loaners, and handling quality holds. When evidence is ambiguous, the system escalates instead of guessing.
How we built it
We built the project in Python using the challenge datasets and policy handbook. A graph-style relationship index connects information across tables so the system can follow dependencies and explain why an order or vehicle is affected. We separated language-model interpretation from deterministic entity resolution and policy enforcement, keeping operational decisions grounded in validated evidence rather than generated text alone.
What we learned
One of our biggest lessons was that a plausible response is not necessarily a correct response. Date boundaries, shared capacity, customer priorities, and conflicting information can all change the right decision. We learned to validate evidence before proposing actions and to distinguish offline test success from live-model accuracy.
Challenges we faced
Compound incidents required careful resource tracking, while independent events needed isolated state. We also tested handling of rumors, ambiguous names, unsupported quotations, and prompt-injection attempts.
Live integration was another challenge. We implemented an OpenRouter adapter for a free, open-weight model, with structured-output validation, restricted provider routing, and no paid fallback. After fixing test-harness issues, our diagnostic confirmed that the selected provider was temporarily rate-limited upstream. Live extraction accuracy therefore remains unverified.
Accomplishments and next steps
The current prototype passes 64 offline tests and supports dry-run planning through an in-process coordinator and three deterministic specialists. Our main accomplishment is an explainable planning foundation that checks evidence and policy before proposing operational changes.
Next, we want to validate live extraction, build the control-tower interface, and integrate controlled enterprise action execution. The current prototype does not make live operational changes.
