ShubhamBuchwani/META-Hackathon

FinOps Cloud Optimizer — OpenEnv Environment

A production-grade RL environment for cloud infrastructure cost optimization. Evaluates AI agents on multi-step reasoning, numerical analysis, constraint-aware planning, and safety-first decision-making.

Problem

Cloud waste costs enterprises $17.6B/year (Flexera 2024). Traditional FinOps tools are rule-based and miss context-dependent decisions. This environment challenges AI agents to perform real-world FinOps:

1. 1.Analyze resource utilization (average vs. p95 traps)
2. 2.Identify waste across multi-account, multi-region infrastructure
3. 3.Propose safe optimizations (rightsize, terminate, schedule, reserve, migrate)
4. 4.Maintain production SLAs — breaking prod is heavily penalized

No existing RL/AI benchmark covers this domain.

Why It Matters

• —Direct dollar-value outcomes — not proxy metrics
• —Tests investigation-first reasoning — agents must ANALYZE before acting
• —Safety-aware optimization — production incidents reduce score by 40% each
• —Multi-step planning — 15–60 step episodes with dependency chains and traps
• —Real-world deployable — companies could use this to evaluate their FinOps AI tools today

Environment API

The environment runs as a FastAPI server:

Action Space

Observation Space

Partial observability: Detailed utilization metrics (p95, memory, IOPS) for each resource are hidden until the agent calls ANALYZE on that resource.

Tasks

Task 1 — Sandbox Cleanup (Easy)

• —Scope: 1 AWS account, 1 region, 8 resources
• —Spend: $4,200/month | Target: -15% ($630)
• —Optimal savings: $890
• —Max steps: 15
• —Trap: Staging EC2 has low average CPU but high p95 from CI/CD — don't terminate it, rightsize it.

Task 2 — Cross-Account Right-Sizing (Medium)

• —Scope: 3 AWS accounts, 2 regions, 20 resources
• —Spend: $28,500/month | Target: -20% ($5,700)
• —Optimal savings: $7,200
• —Max steps: 30
• —Traps: (1) Production RDS looks idle on average but p95 is 78% during nightly batch. (2) Two reserved instances expire in 30 days — proactively identify and renew.

Task 3 — Enterprise Transformation (Hard)

• —Scope: 5 AWS accounts, 3 regions, 45 resources
• —Spend: $142,000/month | Target: -25% ($35,500)
• —Optimal savings: $48,000
• —Max steps: 60
• —Traps: (1) Mislabeled "development" account serves production tools. (2) GPU instances need scheduling, not termination. (3) NAT Gateway can be replaced with VPC endpoints. (4) Circular dependency between cross-account resources.

Reward Design

Dense per-step reward:

R = R_savings + R_investigation + R_safety + R_efficiency

Grading

Each episode is scored 0.0–1.0:

Score = 0.40×Coverage + 0.30×Safety + 0.15×Efficiency + 0.15×Diligence

Task-specific bonus modifiers reward proactive optimization (RI renewal, architectural flags, trap avoidance).

Setup

Prerequisites

• —Python 3.11+
• —Docker (for containerized deployment)

Local Setup

git clone <repo>
cd finops-env
pip install -r requirements.txt
uvicorn main:app --host 0.0.0.0 --port 7860

Docker

docker build -t finops-env .
docker run -p 7860:7860 finops-env

Running the Baseline Agent

export API_BASE_URL=https://api.openai.com/v1
export OPENAI_API_KEY=sk-...
export MODEL_NAME=gpt-4o-mini
export ENV_URL=http://localhost:7860

python run_baseline.py --task sandbox_cleanup
python run_baseline.py --task cross_account_rightsizing
python run_baseline.py --task enterprise_transformation

Baseline Results

Hugging Face Space Deployment

Set environment variables in Space settings:

• —API_BASE_URL — LLM inference endpoint
• —MODEL_NAME — Model identifier
• —HF_TOKEN — HuggingFace token
• —OPENENV_TASK — Task to run

Space type: Docker | Port: 7860

Validation

# Validate openenv.yaml
openenv validate openenv.yaml

# Quick smoke test
curl http://localhost:7860/health
curl -X POST http://localhost:7860/reset -H "Content-Type: application/json" -d '{"task_id":"sandbox_cleanup"}'
curl -X POST http://localhost:7860/step -H "Content-Type: application/json" -d '{"action_type":"analyze","target_resource":"ec2-sandbox-dev-01","parameters":{}}'
curl http://localhost:7860/grade

• —All 3 tasks run in < 3.5 minutes total
• —Environment memory: < 160MB RAM (pure Python, no ML models)
• —Runs comfortably on 2 vCPU / 8GB RAM