Inspiration
As Masters students in Business Analytics/Computer Science, we recognized a "dirty secret" in our field: GenAI is an essential tool for the future of work, but it hides a staggering carbon debt. In a world where simple queries like "what is the weather today?" get directed to the biggest, most power-consuming model, the human reliance on AI might cost us Mother Earth. We wanted to build a system where using AI at a larger scale could be made more sustainable.
What it does
GenAI is an essential tool for the future of work, but its huge power requirements have a massive carbon footprint. GreenRouting is a "brain" that sits between you and the AI to fix this. It works by checking how "clean" your local power grid is in real-time.
If the grid is currently relying on "dirty" energy like coal, the app automatically routes your request to a smaller, eco-efficient model like Amazon Nova Lite. If the grid is clean and your task is complex, it scales up to a high-power model like Nova Pro. This ensures you get the intelligence you need without causing unnecessary harm to the planet.
How we built it
We developed a Serverless Hub-and-Spoke architecture using Python and Streamlit.
1. The Intelligence Layer: We built a GreenRouter that analyzes prompt complexity and multimodal attachments.
*2. The Environmental Layer: * A Carbon API Observer fetches real-time grid intensity ($$gCO₂/kWh$$) to determine the "health" of the energy grid.
3. The Inference Layer: We integrated Amazon Bedrock to access the Amazon Nova family, allowing us to route between "Eco-Paths" (Nova Lite, Llama, Claude, Mistral) and "Power-Paths" (Nova Pro, Claude Sonnet, Llama 70B, Mistral Large).
Challenges we ran into
Our biggest challenge was ensuring reliability without sacrificing our mission. We solved this by implementing Self-Healing Inference, which re-routes a task to a "Pro" model if the "Lite" model’s output is insufficient. We also had to manage the complexity of real-time data flow from multiple AWS services like DynamoDB and Bedrock while maintaining a minimalist, responsive UI.
Accomplishments that we're proud of
Through the implementation of Semantic Response Caching, we discovered that identifying repeated queries using rapidfuzz could achieve 100% energy savings on inference, proving that the greenest prompt is the one that never has to be run.
What we learned
Building GreenRouting taught us that the "Carbon Debt" of AI isn't just a hardware problem—it’s a software design problem. We learned that transparency is the first step toward change; by implementing the SCI specification, we moved from vague "green" claims to measurable, mathematical proofs of impact. We realized that technical savvy users at places like Amazon don't just want efficiency; they want quantifiable auditability.
Perhaps most importantly, we learned that intelligence is a spectrum. Not every task requires a high-reasoning, energy-intensive model like Nova Pro. By building a "brain" that assesses complexity before compute, we discovered that we could drastically reduce emissions without sacrificing the user experience. We learned that an agentic approach can turn a company's massive operational scale into a tool for environmental good.
What's next for GreenRouting
While GreenRouting is currently a standalone platform, our next step is to scale and integrate it directly into existing enterprise systems. We are targeting companies that run massive agentic operations dispatching hundreds or thousands of prompts per minute, to whom our routing logic can provide compounding costs and carbon savings.
On the technical roadmap, we plan to: 1. Expand the Multi-Agent Registry: Integrate a wider variety of specialized foundation models from Amazon Bedrock to provide even more granular routing options.
2. Enhance ESG Reporting: Deepen our SCI (Software Carbon Intensity) scoring to include more variables, such as network transit emissions, to provide an even more accurate "Carbon Profit" proof for the global carbon trading market.
3. Global Grid Monitoring: Scale our Carbon API Observer to monitor data center grids worldwide, allowing international corporations to optimize their global AI footprint through a single dashboard.
Built With
- amazon-web-services
- aws-bedrock
- aws-cloud
- aws-dynamodb
- aws-iam-api-electricitymaps-(for-carbon-intensity-values)-python-libraries:-streamlit
- awsbedrock
- awsdynamodb
- awsiam
- boto3
- css
- css-html-platforms-and-services-github-repository
- electricitymapsapi
- github
- html
- pandas
- plotly
- python
- python-dotenv
- rapidfuzz
- reportlab
- requests
- streamlit
Log in or sign up for Devpost to join the conversation.