Sinegas23/Metaxscaler

Cloud Infrastructure Cost Optimizer OpenEnv

An AI-powered environment for autonomous cloud cost management (FinOps). Agents learn to navigate complex cloud dashboards, identify idling resources, and execute optimization strategies while maintaining strict performance SLAs.

🌟 Motivation & Real-World Utility (30% Evaluation)

The Problem: Enterprises waste over $10 Billion annually on over-provisioned cloud infrastructure. Traditional "static" rules fail because they cannot balance the tradeoff between cost savings and performance reliability.

The Solution: This environment fills a critical gap in the OpenEnv ecosystem by providing a high-fidelity FinOps benchmark. Agents are trained not just to "cut costs," but to act as Sovereign Cloud Engineers who must weigh the financial impact against production stability.

• —Immediate Value: Evaluating how LLMs reason about resource utilization (CPU/Memory) vs. cost centers.
• —Novelty: Unlike simple chat or moderation tasks, this requires precise mathematical reasoning about capacity planning.

🛠️ Action Space (Discrete Multi-Field)

👁️ Observation Space (Structured JSON)

The agent monitors a dashboard of live Resources:

• —resource_id: The identifier.
• —cpu_utilization: 7-day average utilization %.
• —monthly_cost: Current USD run rate.
• —is_production: Flag for critical SLA resources.

🎯 Task Tiers & Difficulty

💰 Reward Design (Partial Progress & Safety)

• —Cost Savings: +Reward for every USD saved monthly.
• —Partial Progress: Immediate positive signals for each resource optimized in a cluster.
• —Safety Penalty: -1.0 penalty for "Destructive Actions" like stopping a Production node unexpectedly.
• —Efficiency Bonus: Extra points for achieving the most optimized SKU recommended by the Oracle.

🚀 Setup & Usage

Running Locally

1. 1.pip install -r requirements.txt
2. 2.python main.py
3. 3.Verify via BASELINE_MODE=mock python inference.py

Docker Deployment

docker build -t cloud-optimizer .
docker run -p 7860:7860 cloud-optimizer

📊 Baseline Scores (Reproduce via mock mode)