GreenRoute Nemotron Audit Lab

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for GreenRoute Nemotron Audit Lab

Inspiration

AI products often send every request to the same powerful model, even when a smaller model could meet the quality requirement. That can increase cost and estimated energy use while hiding where the request is processed. GreenRoute Nemotron Audit Lab is intended to make those routing decisions inspectable and testable before deployment.

Disclosed baseline

The linked GreenRoute Audit Lab is an existing standalone browser prototype created before this Nebius competition. The user describes an AI workload, chooses a required-quality setting, and applies a data-region policy. A deterministic local policy engine classifies the workload, filters a demonstration model catalogue, selects the lightest adequate route, and compares estimated energy and CO2e ranges with a heavy-model baseline. Every decision can be exported as prompt-free JSON audit evidence.

The baseline uses semantic HTML, responsive CSS, vanilla JavaScript, and a versioned demonstration catalogue. Prompt text is neither transmitted nor stored. The interface labels impact figures as decision-support estimates with visible uncertainty rather than meter readings or verified avoided emissions.

New competition work

For this hackathon, the planned new work is a Nebius-hosted audit extension using an eligible NVIDIA open-source model. The model will help evaluate route explanations and failure cases while the deterministic policy and human-review boundary remain visible. The runtime, model identity, prompts, outputs, and evaluation traces must be documented and reproducible.

A real Nebius Token Factory or AI Cloud deployment and an eligible NVIDIA model are not claimed as implemented yet. They will only be claimed after the Builder Program email is verified, free competition credits are available without a payment card or billing commitment, and the repository and live demo prove the integration.

Responsible measurement

The central challenge is avoiding false precision. Provider hardware, batching, live grid mix, and per-query electricity use are usually unknown. The audit therefore preserves assumptions, uncertainty ranges, quality guardrails, rejected alternatives, and human approval. Sustainability never silently overrides quality or data-region requirements.

Planned evaluation

• Compare route quality against a documented heavy-model baseline.
• Test incorrect classifications, unavailable regions, and insufficient-capability cases.
• Record when the NVIDIA model agrees, disagrees, or cannot verify a policy explanation.
• Export prompt-free audit evidence and a separate model-evaluation trace.
• Document limitations, costs, and the exact Nebius/NVIDIA runtime configuration.

Next steps

1. Verify the Nebius Builder Program email sent to mail@dinentreprise.dk.
2. Confirm free competition credits without adding a card or starting a paid plan.
3. Implement the eligible NVIDIA model on Nebius.
4. Publish the competition-specific code, license, README, evaluation results, hosted demo, and a new public video under three minutes.

Build statement

The linked GreenRoute repository and demo are the disclosed pre-competition baseline. Nebius and NVIDIA integration are planned competition work and are not presented as complete until independently visible in the code, runtime, evidence, and new demo.

Built With

• json