ChainPilot

Inspiration

Supply chains power everything we buy, yet companies lose up to 30% of their operating costs to inefficiencies they never even notice. The problem is not that people are not trying, it is that no human can watch every supplier, warehouse, and route around the clock. We built ChainPilot because AI agents can.

What it does

ChainPilot deploys a team of AI agents that never sleep: one hunts for inefficiencies the moment they appear, another validates that the fix is safe, and a third executes the change before a human even notices something went wrong. It is like giving your supply chain its own nervous system that feels the pain, diagnoses the problem, and heals itself in real time.

How we built it

Built as a multi agent orchestration system powered by NVIDIA Nemotron
Deployed on NVIDIA Brev cloud GPU instances for scalable, high performance inference
Leverages Nemotron's strengths in advanced reasoning, function calling, and autonomous decision making
Three specialized agents coordinate through ReAct pattern workflows (reason, act, observe loops)
Red Agent: autonomous reasoning to detect supply chain inefficiencies
Blue Agent: multi step problem solving to validate fixes against policies and constraints
Executing Agent: tool calling to interact with external APIs and execute approved changes
End to end multi step workflows that plan and execute complex optimizations in real time
True agentic behavior that goes beyond chatbot responses into autonomous action beyond chatbot responses.

Challenges we ran into

We learned the hard way that ambition can be your enemy at a hackathon, initially over engineering the system with too many moving parts before stripping it back and focusing on making the core multi agent loop work cleanly end to end.

Accomplishments that we're proud of

We walked in having never touched NVIDIA Nemotron or Brev cloud instances and walked out 24 hours later with three autonomous AI agents reasoning, validating, and executing supply chain optimizations on their own with real time updates streaming to a live dashboard. What makes us proudest is how our team came together, splitting across backend, frontend, and agent architecture, and stitching it all into a fully working product.

What we learned

We learned that Agentic AI is strongest when it is grounded in a real system, not just generating recommendations. Building ChainPilot taught us how to connect agents to live data, constrains, backsend actions and measurable outcomes. We also learned that trust is just as important as automation, so every decision needs to show its reasoning, tradeoffs, and before-and-after impact.