EcoPet AI

Inspiration

We were inspired by the rapid, accelerating use of GenAI models from millions of consumers and the environmental challenges and human impact that current models present to our tech ecosystem. Training a large language model (LLM) comparable to GPT‑3 has been estimated to use around 1,200–1,300 megawatt-hours of electricity and emit roughly 500–600 tonnes of CO₂, comparable to the lifetime emissions of several average cars or hundreds of long‑haul flights. Studies and scenario analyses suggest that, if current growth continues, AI‑driven data centers could emit on the order of 2.5 billion tonnes of CO₂ annually by 2030, roughly 40% of current annual U.S. emissions, absent aggressive decarbonization and efficiency gains. In addition, communities near large AI driven data centers face increased air pollution and health risks where added electricity demand is met by fossil‑fuel plants, reinforcing existing environmental injustices.

We wanted to create an application that seeks to address these challenges in how consumers are using GenAI products, provide an alternative mode to using GenAI models with the environmental impact apparent to user interface, and seek to drive change in user behavior by incentivizing, educating, and making people curious about what the impact and footprint of their queries and searches that are often hidden in GenAI platform user interfaces.

What it does

This tool allows users to input a prompt intended for a GenAI model, then recommends the most energy-efficient LLM models from the top 50 based on the category of their task. Our application shows users the CO₂ cost of different model choices, calculated through machine-learning data prediction score. This hopes to help them understand and be aware of more sustainable and efficient options. To incentivize better choices, it features a virtual pet which grows more energetic as you conserve CO₂ through efficient model and prompt usage, and the energy you save can be spent to unlock cute outfits and accessories—while wasteful or excessive usage makes the pet tired.

How we built it

Data prediction & Calculating Co2 score: we used HuggingFace and scraped top 50 ranked LLM models for overall tasks. Then we trained XGBoost prediction model on data we found on token/inference speed and data center locations for each of the companies to create a CO2 emissions score.

FastAPI: Task classification and recommendation algorithm; endpoints

React + Vite + Tailwind: UI components; Vercel for frontend deployment

Challenges we ran into

Main challenge was not finding enough research or publicly available data around LLM model efficiency based on prompt use and a direct connection to environmental impact. This pushed us to use machine learning model to predict what the estimated grams of CO2 generated by different requests.

Accomplishments that we're proud of

Deployed a production app with SSL at ecopet.boston - our first real domain!
Full stack application development in short amount of time
Applying ML data prediction to a real life problem
Incentivizing product design and user intervention
Potential for real, tangible impact; if 1,000 users save 10g CO₂ daily = 3.65 tons/year!

What we learned

Many things!

Using ML data prediction model pipeline
API deployment
Better git management and version control
Full stack integration and product management

What's next for EcoPet AI

Creating a browser extension for more integration in how people use in their day-to-day.
User Leaderboard and local/global ranking of how much each user is saving in co2 based use of app; team challenges, share your pet's outfits; community focus
Integration and support for local LLM download for users -- would further decrease user carbon footprint
Image/text/video support
Unlock different creatures based on total CO₂ saved
Partnerships where we convert saved CO₂ into real tree-planting credits - so exciting!
API for developers so other can apps integrate our recommendation engine
More accurate and specific CO₂ measures for each prompt employing deeper research findings and building datasets for this use-case.