CarbonSight

agentic structure (adk) demo
embeddings using adk (gemini embeddings model)
adk parent and child agents
model energy network visual
green flag (optimal model chosen)
features/landing page
Auto-Switcher for efficiency

Inspiration

Our team came together from very different motivations but found common ground in sustainability:

Ira was inspired by her first-year project optimizing shell scripts to save water in a code editor.
Advita (me) wanted to explore how to optimize large language model (LLM) usage and cut down their growing carbon footprint — something closely tied to my enterprise’s long-term goal of energy efficiency in homes.
Aarav, a full-stack developer, explores AI/ML applications and builds data-driven solutions
Devank, with strong full-stack skills, was passionate about crafting dashboards and visualizations that make sustainability visible.

Together, we realized enterprises lack visibility into how much energy their AI systems consume. Teams and managers alike are flying blind when it comes to carbon costs. We decided to build a system that not only optimizes model usage but also incentivizes greener choices through transparent blockchain rewards.

What it does

CarbonSight is a carbon-aware optimization framework for generative AI:

Smart Model Routing: Prompts are routed to Gemini models (Flash, Pro, Flash Lite) based on energy thresholds and user-set quality tolerance.
Embeddings & Caching: Similar prompts are detected via Gemini embeddings to avoid redundant compute.
Thinking Budgets: Allocate dynamic reasoning tokens for complex tasks vs lightweight queries.
Employee Experience: Inline feedback after every query → 🔴 “High Energy Model” or 🔵 “Efficient Model.”
Enterprise Dashboards:
Team leaderboards for green usage.
Org-wide statistics for managers.
Exportable blockchain-verified carbon credit reports.
Incentivization: Users earn NFT badges + $GREEN tokens for every “green swap,” pooled transparently on-chain.

How we built it

Google ADK + Gemini: Core agentic framework (routing agent, embedding agent, thinking agent).
FastAPI Backend: Interfaces with ADK and exposes REST endpoints.
React Frontend: Custom chatbot replacing ADK’s default UI, with badges, toggles, and dashboards.
Blockchain: OpenZeppelin ERC-20 & ERC-721 contracts for $GREEN tokens + Proof-of-Green NFTs.
Caching Layer: In-memory semantic similarity cache for embeddings.

Challenges we ran into

Debugging ADK schema mismatches (newMessage.parts.text) that caused 422 errors.
Getting frontend ↔ backend to communicate (CORS + endpoint mismatches).
Defining energy + CO₂ metrics since APIs don’t expose raw usage — we mapped to latency, token counts, and estimated kWh.
Integrating blockchain rewards in real time without hurting latency

Accomplishments that we're proud of

Built a working agentic framework that routes queries across Gemini models dynamically.
Created a React chatbot that displays not just answers but carbon stats + efficiency badges.
Designed a dual-level dashboard system for both employees and managers.
Implemented semantic caching so repeated queries don’t waste compute.
Introduced a novel incentivization strategy: gamified leaderboards + blockchain-backed rewards.

What we learned

How to extend Google ADK for custom agent workflows.
How embeddings + cosine similarity can meaningfully reduce redundant LLM calls.
How to balance output quality vs energy use with thinking budgets.
That enterprises value auditability as much as sustainability- blockchain made this possible.
The power of gamification in driving sustainable behaviors.

What's next for CarbonSight

Production-grade blockchain: Move from testnet tokens to enterprise-ready compliance credits.
Integrations: Slack/Teams bots so employees use CarbonSight directly in their workflows.
Analytics at scale: Plug into enterprise data lakes for org-wide energy forecasting.
Partnerships: With enterprises aiming for Net Zero targets to pilot CarbonSight in real teams.
Open sustainability standard: Advocate for energy usage reporting APIs from model providers.