Demo: https://uclhack-x-thermotrace.vercel.app/
Inspiration Cooling systems in buildings and data centers use a huge amount of energy. A fifteen percent cut in a one gigawatt site can power one hundred twenty five thousand homes. These systems still need experts all day and all night. Existing autonomous tools feel like black boxes. We wanted a system that explains itself and stays safe from the first day.
What it does The agent makes cooling decisions every fifteen minutes. It reads real time data, runs forecasts with high accuracy, and produces clear reasoning for each action. It blocks unsafe choices through constraint checks. ThermoTrace shows predictions, efficiency, and explanations so operators know exactly what the agent is doing and checks physics models to evaluate AI decision without Human.
How we built it We designed a prompt with HVAC knowledge and safety rules. We added a load forecaster and a physics model. The system runs in a simulation that matches real chiller behavior. All decisions pass through a physics layer that checks limits before execution.
What we learned LLMs can manage physical systems when they use the right tools. Clear reasoning builds trust. Physics checks keep the agent safe.
Challenges Hallucination, speed, and model accuracy. Tool validation, prompt tuning, and feature work solved most of it.
What is next Deploy on a real site, train on operator data, scale to data centers and district cooling, and extend to multi agent control.
Built With
- langchain
- langgraph
- langsmith
- python
- react
- typescript
Log in or sign up for Devpost to join the conversation.