Minima

Inspiration

Seeing Preet's Loomin win at the NVIDIA GTC hackathon — a physics AI tutor built for students — and CarbonSense targeting enterprise sustainability made me wonder: what about the environmental cost of AI itself, at the student level? Minima is the intersection: an AI sustainability tutor that doesn't just teach efficiently, it prevents the waste of teaching inefficiently.

What it does

Minima is an autonomous Nemotron-powered agent that intercepts wasteful AI prompts before they're sent. It analyzes scope, estimates projected energy cost in watt-hours, rewrites the prompt sustainably, and only calls Nemotron after the user approves the optimized version. Savings of up to 97% per prompt.

How we built it

• NVIDIA Nemotron-3-nano-30b-a3b via NVIDIA NIM endpoints
• 7-step autonomous agent pipeline: analyze, intercept, count tokens, estimate footprint, diagnose waste, optimize, save memory
• Persistent memory tracking prompt habits over time
• Live policy audit log showing every agent decision in real time
• Deployed on NVIDIA Brev cloud (NemoClaw launchable)

Challenges

Getting the waste detection right — the model sometimes returned unusable rewrites for very broad prompts. The fix was adding retry logic and a deterministic fallback so the agent never crashes.

What I learned

That the hardest part of sustainability is interception, not measurement. Building an agent that stops waste before it happens — rather than reporting it after — required rethinking the entire prompt-to-answer pipeline from scratch.

Built With

• css
• html
• javascript
• node.js
• nvidia-brev-cloud
• nvidia-nemotron-3-nano-30b-a3b
• nvidia-nim-api