Kitsune

Inspiration

We started with a problem visible in our own school: energy use, transport, heating, lighting, and waste all create emissions, but their impact is mostly hidden from students and decision-makers. Awareness alone does not tell a school whether it should fund insulation, LEDs, solar panels, an HVAC upgrade, recycling, or cleaner transport first.

The right choice also changes with the school's budget, building condition, seasonal weather, infrastructure, and operational capacity. We wanted to build a tool that makes these trade-offs visible and lets a school compare possible climate actions before committing real money.

What it does

Kitsune is a school sustainability decision-support system. Its custom environment models four school support hubs across repeated school years. Each hub has a carbon footprint divided into building, lighting, heating, transport, and waste emissions, together with renewable readiness and a retrofit backlog.

The agent can invest in seven physical sustainability actions:

solar panels;
LED lighting;
building insulation;
smart thermostats;
HVAC upgrades;
recycling programs;
electric buses.

Users can change the number of episodes, weeks per episode, starting calendar week, district support budget, sustainability budget, demand pressure, weather intensity, human-error rate, and system-shock rate. Kitsune then reports the current and best carbon footprint, reduction achieved, budget spent and remaining, action frequency, estimated ROI, weather effects, and the state of a school digital twin.

The result is not an automatic purchasing decision. It is a transparent scenario comparison that helps a school administrator or facilities manager understand which combination of actions appears most effective under the selected constraints.

How we built it

We created a custom Gymnasium environment in Python and trained multi-agent PPO policies through PyTorch. Four agents operate school hubs over 52-week school years. Episodes can span multiple years, and annual budgets renew at the start of each simulated year.

The observation space includes operational need, staff capacity, budget health, emissions by school system, renewable readiness, retrofit backlog, season, daily weather pressure, and peer/global indicators. The action space contains 13 decisions, including the seven sustainability investments. Rewards encourage carbon reduction and efficient budget use while also penalizing overspending, unsafe service outcomes, staff burnout, and inequity.

Streamlit provides the interactive control center. It displays training progress, budget allocation, action frequency, CO2 reduction, initiative ROI, daily weather, and the digital twin. Pygame provides an additional simulator view. Qwen is used only in the broader prototype's guarded plain-language support interface; it does not select investments or control the RL policy.

Challenges we ran into

The first challenge was converting a long-term climate problem into weekly decisions. Investments have different costs, delayed benefits, and dependencies, while weather and budgets change what is feasible. We addressed this by separating the footprint into school systems, giving every intervention its own spending and reduction effects, and evaluating policies over complete school years.

The second challenge was reward design. An agent could appear successful by optimizing one easy metric while overspending, ignoring staff capacity, or producing little real reduction. We added counter-metrics and penalties for budget pressure, fatigue, inequity, wrong guidance, missed urgent cases, and low-impact stagnation.

The third challenge was explainability. A final reward number is not useful to a school decision-maker. We therefore expose the selected actions, estimated costs, CO2 reduction, ROI, budget balance, weather conditions, and digital-twin state instead of presenting only a score.

Accomplishments that we're proud of

We built a working custom environment rather than a static carbon calculator. Kitsune can run reproducible scenarios with custom budgets and climate pressure, simulate multiple school years, compare seven sustainability actions, and visualize why a policy behaves the way it does.

We are also proud that the system keeps humans in control. It supports a real investment conversation without claiming that synthetic estimates are guaranteed results or automatically spending a school's budget.

What we learned

We learned that reinforcement learning is useful when decisions are sequential and constrained: an investment made this week changes the options and budget available later. We also learned that a useful optimization system needs counter-metrics. Carbon reduction alone is not enough if the policy overspends, creates inequity, or ignores operational limits.

Most importantly, we learned that explainability is part of the product. Showing assumptions, costs, estimated reductions, and remaining constraints makes an AI recommendation more useful and easier to challenge.

What's next for Kitsune

Our next step is to replace prototype estimates with verified school energy bills, utility factors, building audits, local weather records, vendor quotes, and lifecycle-emission data. We would then add uncertainty ranges, compare the RL policy with simple optimization baselines, and test the dashboard with students, facilities staff, and school administrators.

We also want to support school-specific building profiles, multilingual explanations, saved scenario comparisons, and a pilot in which every recommendation is reviewed by a qualified human before implementation.