GreenInfer

Home Screen
Framework Pipeline
Current LLMs Vs. GreenInfer
Energy Savinges
Energy consumption per model
About page
Impact Page
Framework Page
Chat Page

Inspiration

Every time someone uses ChatGPT, Claude, or any large AI model, they are consuming energy—yet most users are completely unaware.

Through my ISM (Independent Study & Mentorship) research, I discovered that AI is becoming one of the fastest-growing sources of global energy consumption. The root issue isn’t just large models—it’s inefficient usage.

Most prompts contain unnecessary words. For example:

“Can you please help me understand how to write a Python function that sums numbers?” (~15+ tokens) vs. “Write Python function to sum numbers” (~6 tokens)

Both produce the same result, but the first consumes significantly more energy.

At scale, these inefficiencies translate into massive CO₂ emissions.

I wanted to build a tool that makes AI usage more sustainable—without changing how people use it.

What it does

Green-Prompts-Optimizer is an AI-powered web application that makes prompts more efficient and environmentally friendly.

🔧 Core Features Analyzes prompts for inefficiency Optimizes them using a fine-tuned T5 transformer model Calculates energy savings (Wh) and CO₂ reduction (grams) Caches previous optimizations to eliminate redundant computation Tracks personal environmental impact over time 📊 Example Original: “Can you please help me understand how recursion works in programming?” (12 tokens) Optimized: “Explain recursion with examples” (4 tokens) Result: 67% token reduction 0.0008 Wh saved 0.000308 g CO₂ prevented

Multiply this across billions of AI queries, and the impact becomes significant.

How we built it

🧪 Research Phase Analyzed 10+ sources on AI energy consumption Conducted 5 informational interviews with industry professionals Studied prompt efficiency and model behavior Researched optimization strategies used in modern AI systems 📊 Dataset Creation Built 127 high-quality prompt pairs across 6 categories: Technical / Programming Academic / Research Business / Professional Creative / Writing General Knowledge / Lifestyle Technical Support / Education Achieved 72.8% average token reduction 🤖 Model Training Fine-tuned T5-small (60M parameters) 30 epochs, batch size 4, learning rate 3e-4 Used validation split + early stopping to prevent overfitting Trained in ~45 minutes on CPU ⚡ Energy Calculation System Token-based energy estimation: 0.0001 Wh per token (CPU) 0.0003 Wh per token (GPU) CO₂ conversion using U.S. average (0.385 kg CO₂/kWh) Tracks both token-based and time-based consumption 💾 Intelligent Caching Two-tier system: In-memory cache (top 1,000 prompts) SQLite database (persistent storage) Eliminates repeated inference → saves 100% energy on cache hits 🌐 Web Application Flask backend with REST API User authentication + session management Dashboard with: Energy saved CO₂ prevented Token reduction stats Rate limiting to prevent abuse 🧩 Chrome Extension (In Progress) One-click optimization on AI platforms Floating “🌱 Optimize” button Real-time stats syncing

Challenges we ran into

Training Data Quality

Creating high-quality prompt pairs was time-intensive.

Solution: Manually curated diverse examples and validated each for meaning preservation.

Energy Measurement Accuracy

Precise energy tracking requires specialized hardware.

Solution: Used research-backed estimates and token-based modeling inspired by existing tools.

Model Overfitting

Small dataset (127 pairs) risked memorization.

Solution:

Train/validation split (85/15) Early stopping Continuous evaluation

Caching Tradeoffs

Balancing memory usage vs performance.

Solution: Two-tier caching with LRU eviction for efficiency.

Deployment Issues

Model loading errors during cloud deployment.

Solution (in progress): Fixing file paths, dependencies, and model serialization.

Accomplishments that we're proud of

If just 1% of 100M+ weekly AI users adopted this tool:

1M users × 10 prompts/day = 10M optimizations daily 72.8% reduction → 7.28M fewer tokens processed/day Energy saved: ~0.728 kWh/day (~266 kWh/year) CO₂ prevented: ~102 kg/year

That’s roughly equivalent to planting 4–5 trees annually—from a single optimization system.

What we learned

🧠 Technical Skills Fine-tuning transformer models (T5) Energy-aware AI system design Full-stack development (Flask + frontend) Database and caching strategies Chrome extension architecture API design and rate limiting 🌍 Domain Knowledge AI energy usage is a growing global issue Small optimizations scale massively Prompt engineering impacts sustainability Caching is critical for reducing compute waste 💡 Soft Skills Breaking complex systems into components Designing for both usability and impact Building and documenting production-ready systems

What's next for GreenInfer

🚀 Short-term Fix deployment issues Complete Chrome extension Expand dataset to 500+ prompt pairs Support more AI platforms 🌎 Long-term Multi-language support Organization/team dashboards Mobile app (iOS/Android) Integration with AI providers (APIs) Advanced analytics + sustainability reporting Carbon credit tracking + gamification