Sylvo-

Reduce your AI carbon footprint with smart model routing, advanced token pruning, and real-time carbon tracking.

Comment

Sylvo — Carbon-Aware AI Gateway

Inspiration

Every AI prompt has a physical footprint. As AI becomes part of daily life, the energy required to run large models is growing rapidly.

We built Sylvo around a simple belief:

Better AI should not have to mean higher carbon.

Sylvo makes every AI prompt a conscious, carbon-aware choice without changing how people work.

What it does

Sylvo is an intelligent routing layer that manages AI requests based on task complexity and environmental impact.

Key features:

  • Carbon-based routing Simple tasks are routed to small, efficient models (Llama 8B). Complex reasoning is routed to larger models (Llama 70B).

  • Green Boost Using live data from the Electricity Maps API, Sylvo unlocks high-performance models only when the local grid is clean or renewable-heavy.

  • Token pruning Transformers.js removes filler words and unnecessary tokens before prompts are sent, reducing compute and energy use at the source.

  • Live tracking A real-time dashboard shows CO2 saved, energy saved, and token reductions per session.

How we built it

Sylvo is a live web application that acts as a command center for AI usage.

  • Groq API for fast, energy-efficient inference
  • Electricity Maps API for real-time grid carbon intensity

  • A JavaScript Arbitrator that balances:

    • Prompt complexity
    • Model energy cost
    • Live grid cleanliness

Each prompt is routed to the lowest-carbon option that still meets user needs.

Challenges we ran into

Our original goal was a 100 percent local AI system to eliminate network carbon entirely.

In practice, this was difficult:

  • Local inference was slow and inconsistent
  • Setup complexity was high across devices
  • Performance dropped for non-trivial tasks

The pivot

We shifted to smart cloud routing:

  • Cloud models are used only when necessary
  • The smallest, greenest model is always preferred
  • Routing decisions respond to grid carbon instead of being static

This allowed us to ship a working, high-impact system within the hackathon timeline.

Accomplishments we are proud of

The Green Boost feature.

Seeing Sylvo dynamically unlock stronger AI models based on real-time solar and wind availability turns abstract grid data into a feature users can feel and respond to.

What we learned

  • Right-sizing matters — most prompts do not need large models
  • Demand response applies to software, not just hardware
  • Fast pivots unlock impact more than perfect architectures

CO2 and energy calculations

Sylvo provides transparent, real-time calculations for every query and session.

CO2 emissions per query CO2_query = Energy_kWh * GridIntensity_gCO2_per_kWh

Token pruning savings EnergySaved_Wh = TokensPruned * EnergyPerToken * PUE CO2Saved_pruning = (EnergySaved_Wh / 1000) * GridIntensity

Model routing savings CO2Saved_routing = CO2_powerful - CO2_efficient

Session statistics tracked

  • Total tokens processed
  • Tokens pruned
  • Energy saved
  • CO2 saved (routing + pruning)

Forecast and scheduling When the grid is carbon-intensive, Sylvo estimates potential CO2 savings from waiting for a cleaner energy window using live or forecasted grid data.

All calculations are shown directly in the UI with per-query and per-session breakdowns.

What’s next for Sylvo

Next, we are completing the local-first vision with a Chrome routing extension.

Using native browser AI (such as Gemini Nano):

  • Simple tasks run locally by default
  • Cloud models are used only when tasks exceed local capability or the grid is clean

The extension will intercept prompts sent to sites like ChatGPT, evaluate them locally, and only release them to the cloud when justified — making local AI the default and cutting out network carbon wherever possible.

Built With

Share this project:

Updates